GEORGIA DOT RESEARCH PROJECT 22-03
Final Report
STRATEGY ANALYSIS AND EVALUATION FOR EMERGENCY VEHICLE PREEMPTION
AND TRANSIT SIGNAL PRIORITY WITH CONNECTED VEHICLES USING SOFTWARE
IN THE LOOP SIMULATION
Office of Performance-based Management and Research 600 West Peachtree Street NW | Atlanta, GA 30308
September 2023

1. Report No. FHWA-GA-23-2203

2. Government Accession No. N/A

3. Recipient's Catalog No. N/A

4. Title and Subtitle Strategy Analysis and Evaluation for Emergency Vehicle Preemption and Transit Signal Priority with Connected Vehicles using Software in the Loop Simulation
7. Author(s) Angshuman Guin, Ph.D. (https://orcid.org/0000-0001-6949-5126); Somdut Roy, Ph.D. (https://orcid.org/0000-0002-1592-4922); Dickness Kwesiga (https://oricd.org/0000/0002/2642/7825); Gopikrishnan N Suresh Kumar (https://oricd.org/0000/0002/2382/602X) Michael Hunter, Ph.D. (https://orcid.org/0000-0002-0307-9127);

5. Report Date September 2023
6. Performing Organization Code N/A 8. Performing Organization Report No. N/A

9. Performing Organization Name and Address Georgia Tech Research Corporation School of Civil and Environmental Engineering 790 Atlantic Dr. NW, Atlanta, GA 30332 Phone: (404) 385-0569 Email: michael.rodgers@ce.gatech.edu

10. Work Unit No. N/A
11. Contract or Grant No. PI#000253408

12. Sponsoring Agency Name and Address Georgia Department of Transportation Office of Performance-based Management and Research 600 West Peachtree St. NW Atlanta, GA 30308

13. Type of Report and Period Covered Final Report (March 2022 September 2023)
14. Sponsoring Agency Code N/A

15. Supplementary Notes Prepared in cooperation with the U.S. Department of Transportation, Federal Highway Administration.

16. Abstract The overarching objective of this project was to develop and evaluate advanced strategies for Emergency Vehicle Preemption (EVP) and Transit Signal Priority (TSP) implementation that would incorporate and integrate real-time information from connected vehicles, transit vehicles, traffic signal controllers, and other traffic detection technologies to improve overall performance relative to current practice. A microscopic simulation environment was developed in PTV Vissim and calibrated to a real-world corridor for this study. The study demonstrated the benefits of using EVP with dynamic triggering of preemption as compared to no-preemption and traditional preemption using fixed check-in-check-out detectors and developed a novel dynamic preemption logic that minimizes delay for the emergency vehicle while at the same time keeping the preemption period, and hence the disruption to the remaining traffic, to a minimum. The study also developed a machine learning algorithm to generate preemption triggers at lower levels of real-time traffic data availability. The study then created a model to simulate the pull-over behavior of the general traffic, to make the simulation model more realistic and developed code that can be reused by other researchers and practitioners to easily incorporate the advanced model in their simulation. Lastly, the study expanded the effort to TSP and developed some generic guidelines for TSP implementation based on the findings of the study.

17. Keywords Signal Preemption, Signal Priority, Simulation, Software in the Loop

18. Distribution Statement No Restriction

19. Security Classification (of this report) Unclassified

20. Security Classification (of this page) Unclassified

21. No. of Pages 22. Price

320

Free

GDOT Research Project No. 22-03 Final Report
STRATEGY ANALYSIS AND EVALUATION FOR EMERGENCY VEHICLE PREEMPTION AND TRANSIT SIGNAL PRIORITY WITH CONNECTED VEHICLES USING SOFTWARE IN THE LOOP SIMULATION
By Angshuman Guin, Ph.D. Senior Research Scientist
Somdut Roy, Ph.D. Graduate Research Assistant
Dickness Kwesiga Graduate Research Assistant
Gopikrishnan N Suresh Kumar Graduate Research Assistant
Michael Hunter, Ph.D. Professor
Georgia Tech Research Corporation
Contract with Georgia Department of Transportation
In cooperation with U.S. Department of Transportation, Federal Highway Administration
September 2023
The contents of this report reflect the views of the authors, who are responsible for the facts and accuracy of the data presented herein. The contents do not necessarily reflect the official views of the Georgia Department of Transportation or the Federal Highway Administration. This report does not constitute a standard, specification, or regulation.
ii

iii

TABLE OF CONTENTS
EXECUTIVE SUMMARY ............................................................................................ 19 CHAPTER 1. INTRODUCTION .................................................................................. 27
TRAFFIC SIGNAL PREEMPTION AND PRIORITIZATION ......................... 27 BACKGROUND ....................................................................................................... 28 PROJECT PURPOSE .............................................................................................. 29 CHAPTER 2. PROJECT APPROACH........................................................................ 31 PROJECT OBJECTIVES................................................................................................ 31 PROJECT TASKS ......................................................................................................... 32
EVP Activation Strategies .................................................................................. 32 Machine Learning Based EVP Actuation......................................................... 33 Incorporating Driver Pullover Behavior into EVP Strategies........................ 34 Optimal TSP Actuation Strategies .................................................................... 35 CHAPTER 3. SUMMARY RESULTS ......................................................................... 37 OVERVIEW .............................................................................................................. 37 SIMULATION MODEL .......................................................................................... 37 Simulation Environment .................................................................................... 37 Simulation Model and Data Sources ................................................................. 38 Model Runs.......................................................................................................... 39 ERV Behavior...................................................................................................... 40 Model Calibration and Validation .................................................................... 41 EVP Activation Strategies ........................................................................................ 41 Overview .............................................................................................................. 41 Dynamic Preemption (DP) Algorithm .............................................................. 43 Entry and Exit Transitions ................................................................................ 44 Impact of ERV Arrival Time ............................................................................. 45 Dynamic Preemption vs Check-In Check Out ................................................. 47
iv

Exit Transition Analysis ..................................................................................... 49 ML-BASED EVP ACTIVATION STRATEGIES ................................................. 51
Overview .............................................................................................................. 51 Use of Machine Learning to Solve Transportation Problems ........................ 52 ML Model Development..................................................................................... 53 Training and Validation of the ML Model ....................................................... 56 Choice of the Best model .................................................................................... 57 Training the model.............................................................................................. 57 ML Model Evalution........................................................................................... 58 Results .................................................................................................................. 60 INCORPORATION OF DRIVER PULLOVER BEHAVIOR INTO VISSIM ............................................................................................................. 71 Overview .............................................................................................................. 71 Previous Studies of Driver Behavior ................................................................. 73 Observed Driver Behavior in Georgia .............................................................. 74 Modeling Components........................................................................................ 77 Model Parameter Adjustments.......................................................................... 78 Model Design:...................................................................................................... 79 Pullover Frequency ............................................................................................. 80 IMPACT OF SIMULATED DRIVER PULLOVER BEHAVIOR ON EVP STRATEGIES ..................................................................................................... 81 No Pullover versus Pullover models .................................................................. 81 Non-uniformity in Pullover ................................................................................ 84 ANALYSIS OF OPTIMAL TSP STRAEGIES ..................................................... 85 Background ......................................................................................................... 85 Problem Definition and Objectives ................................................................... 87 Methodology ........................................................................................................ 87 Results .................................................................................................................. 94
v

Impact of travel time variability and ETA selection and setting on TSP performance................................................................................................. 104
CHAPTER 4. ANALYSIS AND DISCUSSION......................................................... 114 EMERGENCY VEHICLE PREEMPTION WITH DYNAMIC PREEMPTION LOGIC ................................................................................... 114 EMERGENCY VEHICLE PREEMPTION WITH MACHINE LEARNING LOGIC......................................................................................... 115 IMPACTS OF DRIVER PULLOVER.................................................................. 117 TRANSIT SIGNAL PRIORITIZATION ............................................................. 119
CHAPTER 5. CONCLUSIONS AND RECOMMENDATIONS ............................. 123 CONCLUSIONS ..................................................................................................... 123 STUDY LIMITATIONS ........................................................................................ 126 RECOMMENDATIONS........................................................................................ 129
APPENDIX A. EVALUATION OF EVP ACTIVATION STRATEGIES.............. 135 APPENDIX B. EVALUATION OF MACHINE LEARNING BASED EVP
ACTUATION STRATEGIES ............................................................................... 177 APPENDIX C. VISSIM MODIFICATIONS TO INCORPORATE DRIVER
PULLOVER BEHAVIOR AND COMPARISON OF EVP ACTUATION WITH AND WITHOUT DRIVER PULLOVER ................................................ 233 APPENDIX D. EVALUATION OF OPTIMAL TRANSIT SIGNAL PRIORITIZATION STRATEGIES ..................................................................... 254 REFERENCES.............................................................................................................. 337
vi

LIST OF FIGURES
Figure 1. Photo. Case Study Network of Peachtree Industrial Boulevard: (a) VISSIM Simulation Model, (b) Satellite View by Google MapsTM (Google 2021)............................................................................................................................. 38
Figure 2. Graphs. (a) Impact of staggered ERV entry time on ERV Travel Time, (b) Finding the best cluster number for travel time distribution, (c) Visualization of 3 different clusters separating (ERV entry time, ERV traveltime) data-points........................................................................................................... 46
Figure 3. Graphs. Overall travel time under three entry-transition experimental setups: (1) Preempt Disabled, (2) Check-in Check-out (with Normal Exit), (3) Dynamic Preemption (with Normal Exit) for (a) (top) ERV through the designated route; (b) Non-ERVs at PIB@ Highwoods Center WB-Through for 2 signal cycles after EVP activity................................................................................. 49
Figure 4. Graphs. Mainline approach preemption scenario travel times over the entire mainline (Obenberger & Collura) ERV route: (a) ERVs, (b) Non-ERVs upstream of ERV. ......................................................................................................... 51
Figure 5. Graph. Comparative Results in Violin-plots: ML-Prediction model vs other EVP call experiments on the test-data based on ERV Travel Time (s) .............. 61
Figure 6. Graph. Aggregate Preempt Duration on seven interchanges at PIB for Different Experimental Strategies ................................................................................ 62
Figure 7. Graphs. Non-ERV Travel time Variation with respect to ERV trajectory or Preemption Event: (a) Along main-line PIB, (b) All side-street vehicles since an EVP event....................................................................................................... 63
Figure 8. Graph. Speed variation of Leading and Following Vehicles for ERVs in Tandem Experiment with varying gap between the two ERVs. .................................. 67
Figure 9. Graph. Comparison of ERV KPIs under SILS architecture compared to RBC model based pm ERV travel times. ..................................................................... 68
Figure 10. Graph. Comparison of non-ERV Side-street travel-time after an EVP Call: Variations of EVP call strategy under SILS Architecture ................................... 70
Figure 11. Matrix. Relationship between the EDM and VISSIM..................................... 78 Figure 12. Graph. A comparison of baseline cases without (left) and with (right)
pullover behavior using ERV travel times ................................................................... 82 Figure 13. Graph. Non-ERV travel times observed since the entry of the ERV in
the system without pull-over (Left) and with pull-over (Right)................................... 83 Figure 14. Graph. The ideal/non-realistic pullover behavior compared to the non-
pullover models ............................................................................................................ 84 Figure 15. Matrix. TSP algorithm within VISSIM(R)...................................................... 90 Figure 16. Simulation. Base network model in VISSIM, Far Side Bus Stop ................ 92 Figure 17. Simulation. Two intersection network model ................................................. 92 Figure 18. Graph. SBL delay extent after 10 second EG.................................................. 95 Figure 19. Graphs. Impact on bus travel time at v/c ratio of 0.95 for (a) GE and
(b) EG ........................................................................................................................... 96 Figure 20. Graphs. Impact on bus travel time at three levels of v/c ratio, for (a)
GE and EG and (b) GE only, on Bus travel time cross v/c .......................................... 99
vii

Figure 21. Graphs. Cross Street Delay from GE and EG for (a) SBL and (b) SBT ....... 100 Figure 22. Graphs. Bus travel time vs cycle length for (a) Only TSP affected
buses and (b) All buses............................................................................................... 102 Figure 23. Graphs. Cross street delay vs cycle for (a) SBL and (b) SBT ....................... 103 Figure 24. Graphs. SBT Delay extent, C= 110s for (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 104 Figure 25: Bus travel time variability at v/c=1.0. ........................................................... 105 Figure 26. Graphs. Impact of ETA selection on bus travel time (All Buses) ................. 108 Figure 27. Graphs. Cross street Delay Vs ETA for (a) SBL and (b) SBT...................... 108 Figure 28. Graphs. Bus travel time, Comparing AVL and fixed location check-in. ...... 110 Figure 29. Graphs. Bus Travel time for far and nearside bus stops................................ 112 Figure 30. Graphs. Cross street SBL Delay, Farside vs Nearside bus sto ...................... 113 Figure 31. Photos. Case Study Network of Peachtree Industrial Boulevard: (a)
VISSIM Simulation Model, (b) Satellite View by Google MapsTM (Google, 2021)........................................................................................................................... 141 Figure 32. Graph. Calibration Results: Headway Distribution for Northbound (Obenberger & Collura) Movement at PIB@Medlock bridge Rd: MaxView vs VISSIM ...................................................................................................................... 146 Figure 33. Photos. The Mainline Route chosen for the Study (a) Field ERV GPS data along the mainline overlaid on OpenStreetMapsTM (OpenStreetMap), (b) Mainline ERV Route in Google MapsTM, (c) Mainline ERV Route Static Routing Decision in VISSIM .................................................................................. 150 Figure 34.Graphs. (a) Impact of staggered ERV entry time on ERV Travel Time, (b) Finding the best cluster number for travel time distribution, (c) Visualization of 3 different clusters separating (ERV entry time, ERV traveltime) data-points......................................................................................................... 157 Figure 35. Graphs. Variation in Travel Time for side-street through movement (PIB @ Medlock Bridge Road) for (a) EB Through and (b) WB Through with Different ERV Arrival Times..................................................................................... 159 Figure 36. Graphs. ERV Trajectory in the PIB VISSIM network: preemption enabled: (a) CI-CO, (b) DP ....................................................................................... 161 Figure 37. Graphs. Overall travel time under three entry-transition experimental setups: (1) Preempt Disabled, (2) Check-in Check-out (with Normal Exit), (3) Dynamic Preemption (with Normal Exit) for (a) (top) ERV through the designated route; (b) Non-ERVs at PIB@ Highwo Highwoods Center WBThrough for 2 signal cycles after EVP activity. ......................................................... 163 Figure 38. Graphs. Mainline approach preemption scenario travel times over the entire mainline (Obenberger & Collura) ERV route: (a) ERVs, (b) Non-ERVs upstream of ERV. ....................................................................................................... 165 Figure 39. Graphs. Travel times for side-street trajectories post mainline approach preemption: WB-Through movement at (a) (left) PIB @South Old Peachtree Road, (b) PIB @North Berkeley Lake Road .............................................................. 167 Figure 40. Graphs. Boxplots depicting variations in time taken to travel the entire NB ERV route for Non-ERVs upstream of the ERV: (a) with RBC, (b) with MaxTime i.e., SILS. ................................................................................................ 169
viii

Figure 41. Graphs. Boxplots depicting variations in time taken by ERVs to travel the entire NB ERV route (a) with RBC, (b) SILS. ..................................................... 170
Figure 42. Graphs. Travel time comparison for side-street non-ERVs upstream of ERV, with varying preemption-strategy for EB-through movement on North Berkeley Lake Road at PIB: (a) with RBC, (b) SILS. ............................................... 170
Figure 43. Graph. Average Speed [mph] for ERV at each intersection: Random Seed 1 ......................................................................................................................... 178
Figure 44. Illustration. Schematic Depiction of Factors Determining the need for EVP Call at a given intersection................................................................................. 180
Figure 45. Graph. Trajectory of Other Vehicles around ERV [marked in blue] for one given Random Seed (1) of Simulation where ERV enters the Network at t=3587.5 s [t=0 at 4:30 PM] ....................................................................................... 185
Figure 46. Matrix. Flowchart for Optimal EVP Call using VISSIM simulation iterations. .................................................................................................................... 189
Figure 47. Graphs. PIB@ South Old Peachtree Road: Random Seed 5: (a) Example of ERV trajectory at various EVP call start times; (b) Variation of Stop-bar to Stop-bar travel time, with variation in the start-time of EVP call (trial numbers labelling each data-point).................................................................... 191
Figure 48. Matrix. Flowchart for `Relaxed' Optimal EVP Call using VISSIM simulation iterations. .................................................................................................. 192
Figure 49. Matrix. Flowchart for ML-model Development and Implementation .......... 194 Figure 50. Multiple elements. Schematic Representation of Input Vector of a PIB
intersection ................................................................................................................. 195 Figure 51. Matrix. Schematic Representation of Input Matrix and corresponding
EVP output for a PIB intersection .............................................................................. 196 Figure 52. Graph. Demonstration of Soft-labeling decision-making ............................. 199 Figure 53. Graphs. Demonstration of use of Soft-Labeling Parameters: (a) Case 1:
EVP call is needed; (b) Case 1: EVP Call is not needed and NoPrThres<1 is used............................................................................................................................. 201 Figure 54. Graphs. Validation Set ERV KPI Comparison with Different EVP-MLmodels: (a) Travel-time, (b) Preempt Duration.......................................................... 208 Figure 55. Photo. PIB corridor indicating PIB @ North Berkeley Lake Road being placed significantly downstream (~6500 ft) of PIB @ South Berkeley Lake Road............................................................................................................................ 210 Figure 56. Graphs. Comparative Results in Violin-plots: ML-Prediction model vs other EVP call experiments on the test-data: Mainline-PIB: South Old Peachtree Rd to Howell Ferry Rd: (a) ERV Speed (mph), (b) ERV Travel Time (s) 212 Figure 57. Graph. Aggregate Preempt Duration on seven interchanges at PIB for Different Experimental Strategies .............................................................................. 214 Figure 58. Graphs. Non-ERV Travel time Variation with respect to ERV trajectory or Preemption Event: (a) Along main-line PIB, (b) All side-street vehicles since an EVP event....................................................................................... 215 Figure 59. Graphs. Comparison of ERV KPIs with Changed Simulation Random Seeds: (a) Speed [mph], (b) Travel time [s] ............................................................... 218
ix

Figure 60. Graphs. Comparison of ERV KPIs with Scaled-down Volume: (a) Speed [mph], (b) Travel time [s] ................................................................................ 219
Figure 61. Graphs. Comparison of ERV KPIs with Varying ERV Entry time into the Network: (a) Speed [mph], (b) Travel time [s]..................................................... 220
Figure 62. Graphs. Trajectory of vehicles around and including ERVs in tandem with varying headways between the two ERVs (represented by blue and purple lines): (a) 10s gap (b) 800s gap. ................................................................................. 221
Figure 63. Graph. Speed variation of Leading and Following Vehicles for ERVs in Tandem Experiment with varying gap between the two ERVs. ............................ 222
Figure 64. Photos. ERV Routes: Varying Experimental Strategy on mainline PIB: (a) Original Route: Medlock Bridge Road to Howell Ferry Road, (b) Path 1: Medlock Bridge Road to Pickneyville Park, (c) Path 2: South Berkeley Lake Road to Howell Ferry Road........................................................................................ 224
Figure 65. Graphs. Comparison of ERV Speed with Modified ERV Routes: (a) Path 1: Medlock Bridge Road to Pickneyville Park, (b) Path 2: South Berkeley Lake Road to Howell Ferry Road............................................................................... 225
Figure 66. Graphs. Comparison of ERV Travel Time with Modified ERV Routes: (a) Path 1: Medlock Bridge Road to Pickneyville Park, (b) Path 2: South Berkeley Lake Road to Howell Ferry Road. .............................................................. 225
Figure 67. Graphs. Comparison of ERV KPIs under SILS architecture compared to RBC model: (a) Speed, (b) Travel times. ............................................................... 227
Figure 68. Graph. Comparison of non-ERV Side-street travel-time after an EVP Call: Variations of EVP call strategy under SILS Architecture ................................. 228
Figure 69. Graph. Comparison of VISSIM simulation Runtime, in comparison to wall-clock time, with both VISSIM-RBC and SILS architectures ..................... 230
Figure 70. Matrix. Relationship between the EDM and VISSIM................................... 240 Figure 71. Matrix. Area of effect with indices as reference ........................................... 241 Figure 72. Illustration. Distance limits between vehicles ............................................... 241 Figure 73. Graph. The exponential curve with the cutoff limits highlighted. ................ 245 Figure 74. Graphs. A comparison of three baseline cases without (left) and with
(right) pullover behavior using ERV travel times ...................................................... 247 Figure 75. Graphs. Non-ERV travel times observed since the entry of the ERV in
the system without pull-over (Left) and with pull-over (Right)................................. 248 Figure 76. Graphs. The ideal/non-realistic pullover behavior compared to the non-
pullover models .......................................................................................................... 249 Figure 77. Matrix. TSP Algorithm.................................................................................. 271 Figure 78. Illustration. Base network model in VISSIM, Far Side Bus Stop .............. 273 Figure 79. Illustration. Two intersection network model ............................................... 274 Figure 80. Graph. Arrival profile of buses vs time in cycle ........................................... 277 Figure 81. Graph. Dwell time Distributions from field data........................................... 288 Figure 82. Graph. Dwell time distribution 02 input in VISSIM .................................. 289 Figure 83. Illustration. Nearside bus stop model in VISSIM ...................................... 290 Figure 84. Graph. SBLT delay extent after 10 second EG ............................................. 292 Figure 85. Graphs. Impact on bus travel time at v/c ratio of 0.95 for (a) GE and
(b) EG ......................................................................................................................... 293
x

Figure 86. Graphs. Cross street SBT Delay at v/c ratio of 0.95 for (a) EG and (b) GE............................................................................................................................... 294
Figure 87. Graphs. SBT delay extent with v/c ratio of 0.95 after 10 second (a) GE and (b) EG .................................................................................................................. 295
Figure 88. Graphs. Impact on bus travel time at three levels of v/c ratio, for (a) GE and EG and (b) GE only, on Bus travel time cross v/c ........................................ 297
Figure 89. Graphs. Cross Street Delay from GE and EG for (a) SBL and (b) SBT ....... 298 Figure 90. Graphs. Cross Street Delay from GE only for (a) SBLT and (b) SBT.......... 299 Figure 91. Graphs. SBL Delay Extent, v/c = 1.0 for (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 300 Figure 92. Graphs. SBT Delay Extent, v/c = 1.0, (a) 0 to 10 second truncation and
(b) 10 second truncation only ..................................................................................... 301 Figure 93. Graphs. SBL Delay Extent, v/c = 0.95, (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 302 Figure 94. Graphs. SBT Delay Extent, v/c = 0.95, (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 302 Figure 95. Graphs. SBL Delay Extent, v/c = 0.85, (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 303 Figure 96. Graphs. SBT Delay Extent, v/c = 0.85, (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 303 Figure 97. Graphs. Bus travel time vs cycle length for (a) Only TSP affected
buses and (b) All buses............................................................................................... 305 Figure 98. Graphs. Cross street delay vs cycle for (a) SBL and (b) SBT ....................... 306 Figure 99. Graphs. SBT Delay extent, C= 110s for (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 307 Figure 100. Graphs. SBT Delay extent, C= 130s for (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 308 Figure 101. Graphs. SBT Delay extent, C= 150s for (a) 0 to 10 second truncation
and (b) 10 second truncation only .............................................................................. 308 Figure 102. Graphs. Bus travel time variability at v/c=1.0............................................. 310 Figure 103. Graphs. Bus travel time variability at v/c ratios of 0.95 and 0.85 ............... 311 Figure 104. Graphs. Impact of ETA selection on bus travel time (All Buses) ............... 314 Figure 105. Graphs. Cross street Delay Vs ETA for (a) SBL and (b) SBT.................... 314 Figure 106. Graphs. Bus travel time, Comparing AVL and fixed location check-
in. 316 Figure 107. Graphs. Bus arrival profile at check-in........................................................ 317 Figure 108. Graphs. Bus travel time for different dwell time distributions for (a)
GE & EG and (b) EG only ......................................................................................... 318 Figure 109. Graphs. Bus arrival profile at check-in for DT03 and different offsets ...... 319 Figure 110. Graph. Bus Travel Time for DT03 and different offsets ............................. 320 Figure 111. Graph. Bus Travel time for far and nearside bus stops ............................... 323 Figure 112. Graph. Cross street SBL Delay, Farside vs Nearside bus stop.................... 324
xi

LIST OF TABLES Table 1. TSP strategies and Effectiveness under different levels of demand ................... 98 Table 2. TSP strategy effectiveness for different cycle lengths...................................... 101 Table 3. TSP Effectiveness under different ETA values ................................................ 106 Table 4. TSP strategies and Effectiveness with AVL Algorithm ................................... 109 Table 5. TSP strategies and effectiveness at far and nearside bus stops......................... 111 Table 6. First columns of EVP model results for Validation data (PIB @ North
Berkeley Lake Road) .................................................................................................. 206 Table 7. Summary of the final intersection level ML-models for the seven
contiguous intersections involved in this study.......................................................... 208 Table 8. Summary of MOEs in literature........................................................................ 260 Table 9. Signal Timing Parameters................................................................................. 274 Table 10. Volumes for different v/c ratios...................................................................... 275 Table 11. Splits and maximum GE for varying cycle lengths ........................................ 281 Table 12: Two intersection model volumes.................................................................... 286 Table 13: TSP strategies and Effectiveness under different levels of demand ............... 295 Table 14. TSP strategies at effectiveness at different cycle lengths ............................... 304 Table 15. TSP Effectiveness under different ETA values .............................................. 312 Table 16. TSP strategies and Effectiveness with AVL Algorithm ................................. 315 Table 17. TSP Effectiveness strategies for different dwell time magnitudes ................. 317 Table 18. TSP Strategies for DT03 with different offsets .............................................. 319 Table 19.TSP strategies and effectiveness at far and nearside bus stops........................ 321
xii

LIST OF ABBREVIATIONS

AADT

Annual Average Daily Traffic

AASHTO American Association of State Highway and Transportation Officials

AI

Artificial Intelligence

AIMSUN Advanced Interactive Microscopic Simulator for Urban and NonUrban Networks, Microsimulation Model

APC

Automated Passenger Counts

AR

All Red Signal Phase

ATSPM Automated Traffic Signal Performance Measures

Ave

Avenue

AVL

Automated Vehicle Location

Blvd

Boulevard

BSM

Basic Safety Message

CDF

Cumulative Distribution Function

CI-CO

Check In/Check Out

CNN

Convoluted Neural Network

COM

Component Object Model

CORSIM Corridor Simulation Model

CV

Connected Vehicle

Dr

Drive

xiii

DLL DP E/W EB EBL EBT EDM Eff EG ETA EVP EVPNS ERV FHWA ft G GDOT GE GIS GMM

Dynamically Linked Library Dynamic Preemption East-West Travel Direction Eastbound Eastbound Left Turning Movement Eastbound Through Movement External Driver Module Effective Early Green Estimated Time of Arrival Emergency Vehicle Preemption Emergency Vehicle Preemption Necessity Score Emergency Response Vehicle Federal Highway Administration feet Green Signal Phase Georgia Department of Transportation Green Extension Geographic Information System Gaussian Mixture Model

xiv

GPS GT HILS HGV Hs HSM Hwy IES IQR ITS KL KPI KS KW MARTA ML MOE MPH MUTCD N/S

Global Positioning System Georgia Institute of Technology Hardware in the Loop Simulation Heavy Goods Vehicle Saturation Headway Highway Safety Manual Highway Illuminating Engineering Society Inter-quartile Range Intelligent Transportation Systems Kullback-Leibler Divergence Key Performance Indicators Kolmogorov-Smirnov test Kruskal-Wallis One-way test Metropolitan Atlanta Rapid Transit Authority Machine Learning Measures of Effectiveness Miles Per Hour Manual on Uniform Traffic Control Devices North-South Travel Direction

xv

NBL NBT NoEff NCHRP NTCIP OBU O-D PDF PIB PRC PTV Pkwy RBC RC-Link Rd RITIS R square RSU SBL SBT

Northbound Left Turning Movement Northbound Through Movement Not Effective National Cooperative Highway Research Program National Transportation Communications for ITS Protocol Onboard Unit Origin-Destination Probability Density Function Peachtree Industrial Boulevard Priority Request Server Planung Transport Verkehr, AG, maker of VISSIM Parkway Ring Barrier Controller Roadway Characteristics Database for GDOT Road Regional Integrated Travel Information System Coefficient of Determination Roadside Unit Southbound Left Turning Movement Southbound Through Movement

xvi

SD SILS SpaT SR SSD STM St SUMO SVM TMC TOD TSP US V/C V2I V2V V2X Veh/hr VISSIM VMT

Standard Deviation Software in the Loop Simulation Signal Phase and Timing State Route Stopping Sight Distance Traffic Signal Timing Manual Street Simulation of Urban Mobility Simulation Model Support Vector Machine Traffic Management Center Time-of-Day Transit Signal Prioritization United States Volume to Capacity Ratio Vehicle-to-Infrastructure Vehicle-to-Vehicle Vehicle-to-Anything Vehicles per hour Verhehr in Stdten SIMulationsmodell, Traffic Simulation Model Vehicle Miles Traveled

xvii

VPH

Vehicles per Hour

WATSim Wide-Area Traffic Simulation Model

WB

Westbound

WBL

Westbound Left Turning Movement

WBT

Westbound Through Movement

Y

Yellow Signal Phase

xviii

EXECUTIVE SUMMARY
The overarching objective of this project was to develop and evaluate advanced strategies for EVP and TSP implementation that would incorporate and integrate real-time information from CV, transit vehicles, traffic signal controllers, and other traffic detection technologies to improve overall performance relative to current practice. To ensure maximum fidelity for this evaluation, the project focused on the use of microsimulation methods with a special emphasis on the use of a Software-in-the-Loop (SILS) framework. The key project objectives were:
1. Develop and evaluate strategies for actuation of EVP at traffic signals to minimize travel-times of EVs with a minimal impact on regular traffic,
2. Develop machine learning algorithms to estimate optimal actuation times in the absence of complete real-time traffic state information for EVP,
3. Develop methods for incorporating driver pull-over behavior into the VISSIM modeling framework and evaluate the impact of driver pullover on EVP performance,
4. Develop optimal TSP actuation strategies with special consideration of how late a bus is running relative to its schedule and the potential impact on the remaining traffic using a simulation environment similar to that used for the EVP actuation strategy development.
19

Each of these objectives was treated as a specific research task within the project. A microscopic simulation environment was developed in PTV Vissim and calibrated to a real-world corridor to accomplish these tasks.
This first task of this study demonstrated the benefits of using EVP, combined with CV technology, by developing a DP logic and implementing it in a microscopic simulation model. Dynamic Preemption, with either Normal or In-step exit transitions, led to an approximate 20% (~125 s) reduction in average travel time for the ERV when the ERV traversed through a series of eight preemption-enabled signalized intersections on a highvolume congested corridor. The DP algorithm provided a significantly larger reduction in travel time of the ERV than a traditional check-in-check-out detector-based preemption (~55 s) as compared to the no-preemption case, thus providing a strong demonstration of the advantages of a dynamic approach in preemption that leverages real-time location and traffic data.
The non-ERVs sharing the same path as the ERV also received a significant reduction in delay as a by-product of preemption, primarily due to the queue flush in front of the ERV. When the preemption request was on a mainline approach phase, it was observed that there is a disruption in travel behavior on the side streets, leading to increased travel time with preemption. However, that negative impact on average travel time dissipated rapidly over time. There may also be slight advantages and disadvantages to both the Normal and In-step exit strategies in this regard. For instance, the Normal exit strategy tends to serve the non-ERV traffic side streets slightly better when the preempt call is on
20

the mainline, whereas the In-step strategy tends to serve the non-ERV mainline better when the preempt call was on the side-street.
The primary contribution of this effort, however, is in Task 2, that demonstrates the development of an ML-model that is significantly less data hungry than the Dynamic Preemption model and provides superior ERV performance than either DP or CI-CO. The Dynamic Preemption model requires estimates of the queue lengths at the intersections ahead of the ERV. With heavy penetration of CV technology this would not be a challenge. However, while CV technology penetration is ramping up, the available data are still mostly limited to vehicle detection from traditional infrastructure detectors on the pavement. The ML-model based strategy developed here helps bridge the CV ramp up period gap by providing a methodology that works with the traditional data while it leverages some facets of CV technology that are less onerous to deploy, such as limited deployment of CV units at intersections and on board the emergency vehicles.
One of the major challenges of using ML models is meeting the data requirements for training the ML model. The simulation platform was used to develop and demonstrate the feasibility of creating an ML model for EVP operations. Multiple iterations of simulations were used to generate solutions with varying parameters of simulation and ERV arrivals. The ML-model was trained to learn the characteristics from the available set of solutions with the use of factors such as detector data and signal information. The MLmodel produced results that were better than the performance of the DP solution on the same ERV route.
21

To test for robustness and the potential for generalization of the process in other scenarios and traffic networks, a deeper analysis was performed on the effectiveness of the ML-model. Several key parameters driving the experiment, ranging from simulation random seed, ERV arrival time, variations in ERV route, variations in simulation strategy (e.g. SILS, multiple EVs in tandem, etc.), etc. were tweaked to test the sensitivity of the EVP recommender system. The sensitivity tests provided evidence of the robustness of the model and demonstrated the potential for successful application of the model across different scenarios and networks.
Task 3 of this study addressed the issue that the vast majority of previous studies on preemption have failed to account for the complex interactions between the general traffic and the ERVs. The ones that investigated these interactions typically studied the interactions from an idealized perspective to demonstrate the potential benefits of enhanced V2X communication enabled by CV technology. This study developed realistic models for driving behaviors for interactions between ERVs and non-ERVs which replicate the pullover lane clearing behaviors that are observed in the real world in response to ERVs. The study generated External Driver Model dynamic linked libraries that can be integrated by others for preemption studies to integrate pull-over behavior in the models and provide a realistic baseline that will prevent an unrealistically optimistic bias in the results.
Task 4 of this study expanded the effort to TSP. Overall, it is seen that TSP performance is most favorable in lower v/c conditions where far side bus stops are present. The lower v/c and greater distance between the upstream bus stop and the intersection lessens the uncertainty in ETA, which is critical to TSP effectiveness. Green extension
22

also provides the most benefit to individual buses. However, as congestion increases the effectiveness of TSP decreases. On a highly congested corridor, i.e., v/c ratios approaching or exceeding 1.0, it is possible that TSP may become infeasible as the non-TSP movements may have insufficient slack in available capacity to recover from the TSP related green truncation.
Underlying TSP is a balancing of transit and other vehicle demands. The signal timing for optimal transit vehicle performance may not be optimal for other vehicles in the network. For instance, it was seen that slightly higher cycle lengths or adjusted offsets compared to those for demand based optimal signal settings may result in better TSP performance, as well as lower impacts to non-transit vehicles during TSP events. However, during cycles without TSP the delay experienced by the non-transit vehicle will be higher than under the lower cycle scenario. One implication of this finding is that the setting of a corridor's signal timing parameters should reflect the corridor purpose. That is, if the corridor is designated to serve a significant transit function, then base timing parameters should be selected to improve bus travel time. Where transit is not a primary focus then base signal timings such as higher cycle lengths or TSP based offset may not be warranted.
Based on the conclusions drawn from this study, the following recommendations are made to assist GDOT decisions, on a project level basis, about EVP and TSP deployments decisions and strategies:
The findings of the study clearly demonstrated the advantages of using a dynamic preemption logic over using a fixed check-in-check-out detector based
23

preemption logic. Using a dynamic logic for preemption is therefore recommended under most circumstances where real-time traffic information is available. In cases where accurate queue estimates are not available from the field, ML models can be used to work with conventional vehicle detection data streams, as demonstrated in task 2 of this study.
The study also compared the impacts of two exit strategies for preemption, normal preemption and in-step preemption. It was found that the Normal exit strategy tends to serve the non-ERV traffic side streets slightly better when the preempt call is on the mainline, whereas the In-step strategy tends to serve the non-ERV mainline better when the preempt call was on the side-street. Hence the choice of the exit strategy will need to take into consideration factors such as the demands and turn ratios at the intersections, the use of coordination on the relevant corridor, etc.
The pull-over study helped validate the results of the EVP study in a more realistic simulation scenario. Even when the non-ERVs pulled over for the ERV, the benefits of using EVP in reducing delays for the ERV was still evident and clearly makes the case for use of EVP at signalized intersections.
Based on the findings from the pull-over modeling study an additional argument can be made for deploying EVP to avoid disruption to the traffic along the path of the ERV. It is clear that pull-over causes disruption to the traffic traveling in the same path as the ERV. The comparison of the non-ERV travel times between 24

EVP and non-EVP scenarios, with pull-over integrated in the model, clearly shows that EVP can minimize the disruption with a relatively minor short-term disruption to the cross traffic.
When considering the setting of TSP on a transit designated corridor:
If intersections with a v/c ratio on the order of 0.95 or higher exist, it is recommended to consider slightly longer cycle lengths (on the order of 10 to 30 seconds) to determine if additional slack (unused capacity) in the timing may be obtained. The general traffic delay will need to be checked, comparing the optimal cycle length vs transit cycle length, to determine the acceptability of this option.
Where bus stops exist upstream of an intersection offset that maximize the opportunity for bus passage (include GE time) should be investigated. Intersection dwell time distributions (typically available from the given transit agency) should be included in the modeling of the signal timing.
Where conflicting movement delay is highly sensitive to TSP (typically in higher v/c situations) it may be desirable to limit TSP to GE as GE tends to higher benefits with lower impacts than EG.
While not directly studied in this effort selective TSP may be considered. For instance, only providing TSP to those buses that are running behind schedule. 25

The selection of low (free flow) vs high (congestion based) ETA should be considered in relation to the corridor objectives. While free flow speed-based ETA will likely provide overall better service, where the focus is on congested conditions longer ETA may prove a more suitable option.
Where possible AVL (or other) solutions that allow for flexibility in the selection or application of ETA should be considered.
While beyond the scope of a signal timing only effort consideration should be given, where possible, to the placement of bus stops on the far side rather than near side of an intersection.
The above TSP recommendations represent a generalization of the study result, within the stated limitations. As every corridor has unique characteristics, each corridor should be modeled to determine the most effective application of these recommendations. However, key to the application of any transit timing is that transit timing be considered as part of the signal timing objectives rather than an afterthought to be applied by "tweaking" the general traffic "optimal" results.
26

CHAPTER 1. INTRODUCTION
TRAFFIC SIGNAL PREEMPTION AND PRIORITIZATION
In its broadest sense, traffic control is an engineering approach to implementing the allocation of a scarce resource, the use of a highway segment or intersection, among a variety of potential users in a safe and effective manner. For lower traffic volumes, this allocation can be handled by simple passive measures (e.g., stop and yield signs) but higher traffic volumes and/or complex intersections require active management of roadway access among users, typically implemented through traffic signal systems. In these systems access to "green time" (i.e., permission to execute a particular vehicle or pedestrian movement) is made considering a variety of factors including social costs, "fairness", and the impacts on other elements of the overall network.
Under normal circumstances, all vehicles executing the same movements in an intersection are considered equally. However, there are circumstances where other considerations may make it desirable for certain users to obtain priority in the allocation of "green time" to their movements. For example, we may wish to facilitate the movements of emergency vehicles through the transportation network based on other social priorities. In this case we may wish to override the normal operation of a traffic signal (signal preemption) to allocate green time to the movement of the emergency vehicle. Even small reductions in response times of emergency vehicles have a critical impact in preventing and reducing the loss of life and damage to property. For example,
27

lower EV response times have been shown to result in lower mortality rates when considering significant health emergencies (Okeefe, et al. (2011), Byrne, et al. (2019)).
In other cases, instead of completely overriding normal signal operations we may wish to temporarily increase green time along a particular movement to allow passage of a high occupancy vehicle (e.g., a transit bus) though an intersection that might otherwise be stopped by normal signal operations (signal prioritization). Enabling transit signal priority can help ensure reliable and efficient travel times for transit vehicles that, in turn, has an impact on transit usage and the related traffic demand on our congested roadways.
While these may preemption and priority approaches be worthwhile goals, they may be difficult to implement in practice. For example, it is necessary to understand when a priority vehicle will approach and intersection and, based on current conditions, know when, or if, the current signal cycle should be interrupted.
BACKGROUND
Emergency Vehicle Preemption (EVP) and Transit Signal Prioritization (TSP) are not new ideas and various enabling technologies have been developed and successfully deployed. To date, most implementations of signal preemption/prioritization have involved either special traffic signals (e.g., at the exit to a fire station) or various types of vehicle-mounted systems to let the signal controllers know of the approach of a particular type of high priority vehicle. These latter systems are typically based on either a special detector located a fixed distance away from the intersection or are based on a line-of-sight connection between the vehicle and
28

a special detector. This fixed distance and/or line-of-sight requirement has thus far served to limit the benefits of these technologies (see for example: Nelson and Bullock (2000) or Baker, et al., (2004)). In addition, most of the previous deployments have not directly assimilated real-time corridor traffic demand to optimize performance and thus may also be suboptimal because they focus on a particular intersection rather than on the overall route of the priority vehicle.
With the recent availability of real-time traffic status data from connected vehicles and infrastructure, it is feasible to develop signal control strategies and algorithms that can proactively create a free-flow path through the signalized intersections for the emergency and transit vehicles. While such a methodology has been proposed before and has seen limited implementation using GPS and cellular-phone based technologies, to date there have been only limited efforts at before-after evaluations of distributed predictive EVP and TSP implementations. Research is also limited on the methodology of implementation, especially with consideration of vehicle behavior factors such as crossing over the centerline or use of turn bays by emergency vehicles to traverse an intersection, pull-over of non-emergency vehicles to create a path for emergency vehicles, etc. (Nelson and Bullock (2000)). These existing studies are discussed in Appendix A.
PROJECT PURPOSE
Connected Vehicles (Cvijovic et al.) have been envisioned as providing opportunities for improving traffic safety and mobility in a wide variety of ways including better systems for traffic signal preemption and prioritization. To date, despite a large number of potential
29

applications, there has been only minimal guidance available for their implementation. GDOT has been a nationwide leader in CV deployment and has a strong interest in the development and deployment of advanced traffic management solutions that leverage the high-fidelity data and real-time data driven control made feasible by CV. Improved traffic signal management with Connected Vehicles offers the potential to significantly improve the implementation of EVP and TSP systems by optimizing the overall vehicle pathway through a coordinated signal network. Even before this project, GDOT and local city/county Departments of Transportation have been making significant strides in the planning and deployment of EVP and TSP systems. This project was designed to fill important gaps in knowledge between theoretical proposals and field-ready solutions for implementation of EVP and TSP that can benefit from the realtime information provided by CV infrastructure. The project focuses on evaluating the potential impacts of different EVP and TSP strategies in a simulated environment before field deployment. To ensure a true representation of actual field conditions, the simulations used in this study incorporate Software-in-the-Loop (SILS) traffic signal control by using a software emulation of the same traffic controllers (MaxTime) that are in current use.
30

CHAPTER 2. PROJECT APPROACH
PROJECT OBJECTIVES
The overarching objective of this project was to develop and evaluate advanced strategies for EVP and TSP implementation that would incorporate and integrate real-time information from CV, transit vehicles, traffic signal controllers, and other traffic detection technologies to improve overall performance relative to current practice. To ensure maximum fidelity for this evaluation, the project focused on the use of microsimulation methods with a special emphasis on the use of a Software-in-the-Loop (SILS) framework. The key project objectives were:
1. Develop and evaluate strategies for actuation of EVP at traffic signals to minimize travel-times of EVs with a minimal impact on regular traffic,
2. Develop machine learning algorithms to estimate optimal actuation times in the absence of complete real-time traffic state information for EVP,
3. Develop methods for incorporating driver pull-over behavior into the VISSIM modeling framework and evaluate the impact of driver pullover on EVP performance,
4. Develop optimal TSP actuation strategies with special consideration of how late a bus is running relative to its schedule and the potential impact on the remaining traffic using a simulation environment similar to that used for the EVP actuation strategy development. 31

Each of these objectives was treated as a specific research task within the project. A brief description of the approach used for each objective is given below with a more detailed description provided in the corresponding Appendix to this report.
PROJECT TASKS
EVP Activation Strategies
This task was designed to develop candidate Dynamic Preemption methods for EVP and to compare the effectiveness of these strategies relative to existing fixed-distance EVP actuation as well as a no-activation condition. This evaluation included both the https://www.gti.gatech.edu/consortiummemberseffectiveness in reducing Emergency Vehicle travel times as well as the necessary recovery times for the system to return to equilibrium. This later analysis considered several potential exit strategies.
Previous work in this area as well as details regarding the individual approaches considered are provided in Appendix A. These approaches considered a wide range of conditions expected to influence the effectiveness of EVP under specific conditions. These included various EVP implementation approaches (all-red, fixed-time EVP, predictive EVP) with respect to settings such as trigger time, duration of EVP sequence, method to return to coordination, side-street vs mainline signal control (e.g., use queueflush when EV approaches intersection on mainline but use all-red when EV approaches from cross street), etc. This appendix provides a thorough description of the individual
32

approaches considered as well as extensive results from these simulations. A summary of these results is provided in Chapter 3.
Machine Learning Based EVP Actuation
In the future, CV data, especially Basic Safety Messages (BSMs) from passenger cars, are expected to provide a means to obtain a real-time picture of the traffic state in terms of the queue and the demand that needs to be served at an intersection ahead of an EV or a Bus. However, given the overall slow rate of fleet turnover, the penetration of CV in the general traffic stream is likely to be low for many years. In the meantime, traffic responsive EV and TSP strategies require interim strategies to deal with this limited information regarding queue lengths at upcoming intersections. This project task evaluated the extent to which machine learning methods could develop better estimates of optimal actuation times in real-time, using a limited set of inputs including:
Traffic count estimates from presence detectors at stop bars (typically biased to the underside),
Traffic count from upstream count detectors, Time of day / day of week based typical demands, Time of day / day of week based typical turn ratios at intersections, Injection point (arrival of EV) in the signal cycle, Coordination state and offset, and Any BSM data available (possibly sparse) used to estimate queue lengths.
33

The simulation environments developed in other tasks were used to translate the aggregate level inputs identified here into vehicle trajectories (surrogate for BSMs) that provided the input required to optimize the actuation times. Then these aggregate level inputs and the actuation time outputs were used to train a machine learning model. Once trained, the machine learning model provided recommended actuation times as a function of the aggregate inputs from the field in real-time without the need for the simulation environment in the middle. Detailed information regarding the development of the synthetic-BSMs and simulation results are provided in Appendix B. A summary of important results is provided in Chapter 3.
Incorporating Driver Pullover Behavior into EVP Strategies
The primary modeling system used for the other EVP and TSP tasks relied on the VISSIM microsimulation algorithms to represent the background traffic flow for the Emergency vehicle to navigate. The Dynamic Preemption strategies developed in these tasks aimed to ensure that not only does the Emergency Vehicle (in this case a fire truck) avoid entering the back of a queue at an intersection, it also goes through the intersection without having to slow down. However, this approach had an embedded the assumption that the other vehicles had nowhere else to go but forward through the intersection to allow the passage of the EV. This is a very conservative operational assumption. Nonemergency vehicles are supposed to pull-over to yield a path for the emergency vehicles, if possible. To study the optimal actuation requirements in the case where vehicles could pull-over to allow the Emergency Vehicle to pass, the project developed the necessary
34

modifications to the VISSIM framework necessary to model this non-emergency vehicle behavior in the simulation model, in addition to the queue clearance through the intersection. This task involved implementation of additional vehicle control options in the "COM" interface of VISSIM that allowed both additional vehicle movement options, and various levels of vehicle compliance with these requirements.
The approach used to implement the changes to the VISSIM environment are documented in Appendix C along with results from comparing the performance of the EVP Strategies developed in the first task with and without vehicle pullover. These results are also summarized in Chapter 3.
Optimal TSP Actuation Strategies
Similar to Emergency Vehicle preemption, in transit signal priority (TSP) it is essential to ensure that the demand in front of a bus (queued or otherwise) on the relevant intersection signal approach is served before the bus can traverse the intersection. The timing of the actuation of TSP is critical in ensuring an uninterrupted passage of the bus through an intersection while concurrently ensuring that the impact on the remaining traffic is minimal. Previous pilots of TSP have demonstrated a somewhat muted response of transit vehicle adherence to TSP implementation. This task was designed to develop strategies to adjust the actuation time of TSP to improve the traversal time of the bus through the intersection and thereby improve schedule adherence.
35

To evaluate potential strategies, a simulation model for a notional corridor on a route that is currently experiencing schedule adherence issues was developed and used to investigate the opportunity for improvement by controlling the actuation time of TSP. Earlier actuation, such as when the bus is further upstream (potentially more than one intersection) from the target intersection, and the corresponding responses of the controllers were evaluated along with any impacts on the remaining traffic. Information regarding model and scenario development and the simulation results are presented in Appendix D and summary results are presented in Chapter 3.
36

CHAPTER 3. SUMMARY RESULTS
OVERVIEW As discussed in the previous chapter, the major objectives of this project were to examine various approaches toward the incorporation of CV technologies into the process of granting signal preemption to emergency vehicles (EVP) and adjustments of signal timing to provide priority for transit vehicles (TSP) and to analyze their impacts in a microsimulation environment. In this chapter, we will discuss the microscopic simulation environment developed to accomplish these tasks and provide some of the most important results of the simulations. Each task is discussed at length in the accompanying appendices which provide both additional results and significantly greater detail regarding the simulations specific to each task.
SIMULATION MODEL Simulation Environment
The simulation model developed for this project was developed in PTV VISSIM. For the EVP efforts the model simulates a 6.2 mile stretch along the Peachtree Industrial Boulevard (PIB) corridor from Holcomb Bridge Rd at the south-west to Pleasant Hill Rd on the north-east in Norcross, Georgia. The model includes both the roadways and 25 intersections on and around PIB. The model and the network extents are shown in Figure
37

1. For consistency in naming, the PIB directions of travel are labeled as North / South throughout the length of the corridor. The cross-street approaches are labelled as Eastbound (EB) and Westbound (WB). For this model, the system entities, consisting of the network geometry and the signal-heads, were built based on satellite imagery from OpenStreetMapsTM (OpenStreetMap).
Figure 1. Photo. Case Study Network of Peachtree Industrial Boulevard: (a) VISSIM Simulation Model, (b) Satellite View by Google MapsTM (Google 2021).
Simulation Model and Data Sources For signal control, the model used VISSIM's Ring Barrier Controller (RBC) add-on module to simulate signal controllers and preemption strategies for all "internal" simulations, while Intelight's MaxView Advanced Traffic Management System (ATMS) software (Intelight 2017) was used for all Software-in-the-Loop (SILS) simulations. MaxView is commonly used for signal control applications throughout the State of Georgia.
38

Input data was selected from a per-COVID period to ensure that the traffic and signal plans in the model represented typical traffic operations. Signal control data was based on a non-holiday weekday: Tuesday, October 01, 2019. Signal plan information was obtained directly from the field controllers, representing the active plans. The signals along the corridor are semi-actuated coordinated, with a 160 second cycle.
For simulation model calibration, a comprehensive volume study was not available for this corridor. Thus, for the major and minor road approach volumes, data was assimilated from multiple sources including short-term historical turn-volume count data, counts obtained from post-processing stop-bar presence detector activations, recent traffic studies on the corridor, available Automated Traffic Signal Performance Measures (ATSPM), etc. The signals on this corridor are connected to a central server where the high-resolution signal phase and timing (SPaT) data, as well as detection data, is archived. The archived data had vehicle on-off pulse information corresponding to inductive loop detectors upstream of the stop-bar for the major road through lanes. The pulse data was post-processed to generate vehicle detection data for additional volume calibration. Finally, volume balancing and volume constraint computations based on signal cycle allowance and roadway geometry were used to generate estimates to fill volume gaps or inconsistencies in available data.
Model Runs
The replicate runs used in the study partially automated by using Python 3.7 (Python, 2018) scripts to drive VISSIM using its Component Object Model (COM) interface.
39

For each simulation run the network is initially loaded with 50% of the target volume for the first 15 simulation minutes. Then the volume is raised to 100% for the next 75 simulation minutes. Effectively, out of the total 90-minute runtime, the first 30 minutes are used for model initialization, and only the last 60 minutes are used for collecting data to generate performance metrics corresponding to the PM peak hour (5 pm to 6 pm) except for a limited number of runs associated with Transit Signal Prioritization. In each simulation run, a single preemption event is modeled to ensure complete independence of the results related to each actuation.
ERV Behavior
In addition to its default vehicle types VISSIM enables the incorporation of external vehicle models that reflect operational characteristics unique to a given vehicle type. One such model was used to create an ERV vehicle class, with specific features corresponding to a firetruck. The 3D model and characteristics were obtained from the VISSIM website (PTV, 2021a), and the speed and acceleration characteristics of the vehicle were selected to mimic a typical Heavy Goods Vehicle (HGV) in VISSIM. For this study, an ERV must pass through an intersection during a green, using lanes in the correct direction of travel. That is, the simulated ERV behavior does not allow for running a red or passing through the intersection in the lanes of the opposing traffic.
40

Model Calibration and Validation
Model calibration, i.e., adjusting model parameters to maximize the agreement of the model behavior to field observations (Trucano et al., 2006), is an essential step to ensure that the model accurately represents field conditions. While "fine-tuning" selected key parameters of the model to mimic real traffic in the network is necessary, it is not always possible to match the traffic vehicle-to-vehicle. Nor is such matching desirable as this process may lead to overfitting the model, negatively affecting the robustness, translatability, and generalizability of the results.
In this study, the calibration effort ensured that the model sufficiently reflected the field conditions, considering both mid-block free flow speeds and saturation headways departing a signal. Validation tests with travel-time as the performance metric were performed to confirm the sufficiency of the calibration. Model calibration and validation are described in Appendix A.
EVP ACTIVATION STRATEGIES Overview
This task involved developing and evaluating dynamic preemption EVP strategies in terms of providing the shortest travel-time for an ERV while limiting the delays experienced by the general traffic. In the absence of real-world CV data in the study
41

corridor, the study used simulated data streams from the simulation model to replicate future data streams expected from CV data.
The specific simulations used in this task consisted of evaluating the impacts of an ERV route using preemption on 8 contiguous intersections along PIB specifically focusing on EVP impacts in a coordinated system. These evaluations focused on impacts during the PM peak hour, typically the most congested period for this corridor and considered the impacts on both ERV travel times as well as non-ERV travel times both along the route and conflicting traffic. Specific details on the route and the simulations used in these evaluations are given in Appendix A.
This study did not consider the technical aspects of CV implementation such as the range and accuracy of CV equipment. In the evaluation, the following assumptions were made relating to CV infrastructure:
CVs are present in sufficient numbers to yield a reasonably accurate estimate of the queue lengths.
The path of the ERV is known and the ERV can transmit its position in real time. The signal controllers are connected to a control center that can make signal phase
change requests and signal states are known in real time.
42

Dynamic Preemption (DP) Algorithm
This task examined how the potential how real-time field detection and ERV CV data could be leveraged to improve EVP performance over existing fixed-distance "Check inCheck out" (CI-CO) detector-based systems by developing and testing a Dynamic Preemption (DP) algorithm. In this algorithm it is assumed the ERV is continuously updating its position data. However, other traffic is assumed not to do so (i.e., they are not CVs) allowing for earlier field implementation of the proposed method. In this algorithm, the queue-length on each approach of the ERV path is monitored and known. Based on the queue lengths, the time of the preemption trigger is set to try and ensure that the ERV goes through the intersection without (or with minimal) reduction in speed. At a given approach a sufficient time must be allocated for the vehicles in the queue, as well as those in between the end of the queue and the ERV, to clear prior to the ERV arrival. Some transition time must also be allocated for the signal controller to serve the current phase yellow, red-clearance, and any necessary in-progress pedestrian-walk phases.
The primary assumption for the algorithm is that the number vehicles in the approach queue is same as the number of moving vehicles between the back of the queue and the approaching ERV. While a very rough approximation, this accounts for field data limitations since active counts of moving vehicles on the roadway are unlikely to be available in the short term. In the current algorithm a preemption decision may be made as early as when the ERV begins to approach the upstream intersection for the subject approach. Thus, it is assumed that the ERV route is known at least two intersections in
43

advance. The preempt call time is calculated as follows: If "n" cars are present in the queue, assuming that the headway is 2 seconds and an additional reaction time (i.e., startup lost time) of 4 seconds, the time taken to clear that queue would be (4+ 2*n) seconds. To clear the "n" vehicles between the end of the queue and the ERV, an additional 2*n seconds are required. For the transition from current signal state to the preempt phase, an additional 5 seconds is added. This results in an overall total of (9+4n) seconds for the advance placement of the EVP call prior to the ERV reaching the intersection. Therefore, in a corridor with a free flow speed of "v" ft/sec, the preemption will be triggered by the ERV at a distance of (9+4n)*v feet from the intersection, when the intersection in question has a queue length of n cars.
Entry and Exit Transitions
While the entry transition has a significant impact on the ERV's travel-time, the traveltime of the other vehicles are affected by the exit transition as well as the entry transition. To explore the impact of these transitions on the non-ERVs, two different entrances and two different exit (normal and in-step) transitions were included. Details on these strategies are provided in Appendix (A).
EVP ACTIVATION SIMULATION RESULTS The analysis of the impact of the DP algorithm was based on three primary series of simulation runs (Appendix A). These were:
44

1. The impact of the time of arrival of the ERV during the cycle was explored. This investigation uses the DP algorithm with normal exit.
2. The performance of Dynamic Preemption vs Check-In-Check-Out was compared. The comparison uses normal exit in both cases and included variations in ERV arrival time. As a cross check, the DP results were compared using both VISSIM internal signal control algorithm (RBC) and using the MaxTime signal control software in a Software-in-the-Loop (SILS) environment.
3. A comparison of normal and in-step exit transitions was conducted for the DP model, again considering variations in ERV network arrival time.
Impact of ERV Arrival Time
The PM cycle length for all intersections within the corridor is 160 seconds. Thus, to reflect a cross section of possible arrival times 32 different scenarios are created, with successive five seconds increments for the time of introduction of the ERV into the network. To account for stochastic variability, ten replicate runs per scenario are performed using ten random seeds. Thus, a total of 320 simulation runs (32x10) were conducted in the study of ERV arrival time. For these runs, the DP algorithm was used with normal exit transition.
The result of these variations in ERV arrival times are shown in Figure 2 as a hybrid boxplot. In this and subsequent figures, the red square dots represent the mean of the ten
45

replicate runs with top and bottom of the solid box represent the 75th percentile and 25th percentile (Interquartile Range, (IRQ)) respectively. The black line drawn on the solid box is the median; the "whiskers" around the box span within 1.5*IQR of the box boundaries and the points that lie beyond the whiskers are shown individually.

(a)

(b)

(c)

Figure 2. Graphs. (a) Impact of staggered ERV entry time on ERV Travel Time, (b) Finding the best cluster number for travel time distribution, (c) Visualization of 3 different clusters separating (ERV
entry time, ERV travel-time) data-points
46

In the runs each successive ERV entry time is increased by five seconds to ensure sufficient variability such that arrivals are distributed throughout the cycle. A review on the plot shows significant variation in median travel times and IQR across the cycle. To aid in the analysis the data were segregated based on the clustering techniques proposed by Anderson et al. (Anderson et al., 2019) using an optimum number (in this case 3) of clusters. This procedure is described in Appendix A. After clustering, it is observed that the ERV travel-times resulting with varying ERV entry-time into the network can be roughly broken into two different groups (the green cluster and the violet/yellow clusters in the last image) strongly suggesting that variation in the cycle entry-points has a significant impact on the ERV travel-time distribution. A similar effect was observed on the side-street movements (Appendix A). As a result, all subsequent simulations included a stochastic variation on ERV arrival times.
Dynamic Preemption vs Check-In Check Out
These simulations were designed to examine two potential issues associated with CI-CO EVP activation: 1) if the back of the signal queue extends beyond the check-in detector the ERV does not trigger the preemption call until it advances in the queue to the detector location, and 2) in CI-CO activation the call is set at a fixed distance without consideration of the real-time traffic conditions. Under these circumstances, the DP approach might be expected to improve performance by making the preemption request further upstream thus increasing the likelihood of a successful queue flush prior to the ERV arrival at the intersection box. As with the prior experiment, there are 32 runs of
47

ERV network entry samples for each random seed, thus 320 total runs out of the 10 random seeds. The impact of this difference in preempt methods may be seen in the impact to ERV and non-ERV travel times shown in Figure 3. These results show that on average the CI-CO provides an approximately 73s advantage for ERV travel time over the no-preempt option while DP provides a 142s advantage. All of these travel time differences are statistically significant, based on a non-parametric one-way Kruskal-Wallis H-test at 5% significance level. The effect on side street traffic were comparable, with the average travel time being 107s, 130s and 131s for "no preempt", CI-CO, and DP respectively. Similar trends are observed for other intersections as well. Hence the overall gain in ERV response time with negligible excess side-street traffic delay provides a strong argument for EVP in general, as well as using CV-based activation methods. SILS simulations using the MaxTime controller software produced equivalent results (Appendix A).
(a) 48

(b)
Figure 3. Graphs. Overall travel time under three entry-transition experimental setups: (1) Preempt Disabled, (2) Check-in Check-out (with Normal Exit), (3) Dynamic Preemption (with Normal Exit) for (a) (top) ERV through the designated route; (b) Non-ERVs at PIB@ Highwoods Center WB-
Through for 2 signal cycles after EVP activity.
Exit Transition Analysis
While the negative impacts of EVP are likely to be more pronounced on the side streets, it is possible that negative impacts may also be experienced by vehicles on the preemption approach. Therefore, this experiment considered both the preemption approach and the cross-streets. To avoid the confounding effects of turn-related delays due to mixing of through and turn movements, the measurements in this study are made for vehicles involved only in through movements for both the main-line and cross-streets.
For this study only DP was considered, thus three preemption strategies are tested: nopreemption, DP with normal exit transition, and DP with in-step exit transition. The differences in the "normal" and "in-step" exit strategies are discussed in Appendix A but essentially involve transitioning back either into the normal control sequence or moving
49

directly to the phase that has had the least recent green time. As before, the comparison consisted of 32 ERV network entry times with ten replications for each of the three treatments for a total of 960 (32x10x3) simulation runs. For the subsequent analysis the impact of each strategy is considered over six time-intervals after the ERV first enters the network, where each interval is one cycle length, e.g., 160 seconds, allowing for observation before the preemption call is placed as well as dissipation effects. In the analysis interval aggregations are based on the time an ERV or non-ERV leaves a segment rather than enters. Figure 4 summarizes the travel time variation for the ERV in the network under the different preemption scenarios. The average travel time for nopreemption, preemption enabled with normal exit, and preemption-enabled with in-step exit, are 599s, 457s and 454s, respectively, with an approximately 140 seconds average travel time savings for both preemption exit modes. The improvement for both exitstrategies is statistically significant relative to no preemption, while there is no statistically significant difference between the two exit strategies. Another positive aspect of preemption was a reduction in the variation of travel time as reflected by the reduced IQR, seen in both exit transition strategies.
50

(a)

(b)

Figure 4. Graphs. Mainline approach preemption scenario travel times over the entire mainline (Obenberger & Collura) ERV route: (a) ERVs, (b) Non-ERVs upstream of ERV.

Appendix A includes results for additional analyses for vehicles travelling the same route as the ERV and results from movements along representative cross streets. Both results are in general agreement with those shown here.

ML-BASED EVP ACTIVATION STRATEGIES Overview
The DP algorithm discussed above assumes the real-time availability of queue lengths and uses this information to make decisions regarding when to make the preemption call.
51

Despite the assumption that these data are available, the simulation results from task 1 indicated ERV speeds much lower (~ 40 mph) than the free-flow speed on the corridor, estimated to be at 53 mph showing significant potential for improvement. Some of the potential reasons for this difference are discussed in detail in Appendix B, but the principal factor is the absence of real time information from current field detectors as to the number of vehicles present in each segment of the corridor and thus these decisions must inherently be made with incomplete information. Such decisions are common in both transportation and other fields. In recent years substantial progress has been made in the use of Machine Learning (Mohammadi et al.) methods for assisting such decisions. This task was aimed at evaluating the potential for ML for improving timing decisions associated with EVP applications.
Use of Machine Learning to Solve Transportation Problems
While there have been previous studies on the use ML methods in transportation applications (Appendix B), this task explores the use of these methods to improve EVP decisions. Task 1 demonstrated how integration of real-time traffic flow information from CV technologies (OBU on the ERV and RSUs at the intersections) could feed ad hoc heuristic "Dynamic Preemption" (DP) logic to improve the movement of the ERV through the intersections.
An existent weakness of the DP algorithm is that it assumes significant knowledge regarding queue lengths at the intersections along the ERV path, information that is likely to be difficult to obtain in the short term. In the absence of this information, this task was
52

aimed at examining the extent to which the simulations developed in Task 1 could be used to train a sophisticated ML-model to achieve similar performance to the queueinformation based DP method using only information likely to be available in the nearterm. ML Model Development
Appendix B provides a comprehensive discussion of the technical aspects of the development of the ML model used for this evaluation and thus the ML model development will only be summarized here. A supervised ML approach was chosen for the development of an optimal preemption strategy. The training data set for the ML was limited to data that may be obtained on a reasonably equipped corridor today. That is, vehicle detection information is available in real-time, the ERV has GPS equipment and has connectivity with the traffic signals directly or through a central server, and the traffic signals have remote connectivity and stream signal state (SPaT) data available.
The calibrated corridor model presented for the PIB corridor was used to train the ML model with the simulation results assumed to represent perfect knowledge of corridor conditions. However, the ML training data were limited to the data that would be available in a field environment. The development of the ML method consisted of the following steps:
1) Optimal EVP calls were developed for a number of scenarios utilizing "perfect" information, that is, full knowledge of the number and placement of vehicles based on simulated results. These preemption calls were used for the training data 53

set. The optimal preempt call times were developed using an iterative trial and error method.
2) The motivation for proposing an approach that uses partial, but realistically available data in real-time is that it would mimic the current field situation of limited availability of traffic information. However, it is assumed that even with limited available real-time data, the network infrastructure can provide second-bysecond ERV trajectories using CV technology, as well as SPaT messages providing the signal and detector state. These data were used to train a supervised learning model using the solution suggested in step (1) to build a preemption decision algorithm utilizing the ERV trajectory and SPaT messages. Different supervised learning algorithms are tested at each intersection and the best performing models were chosen.
3) The EVP recommender models formed for contiguous intersections were combined and run concurrently to create a system-level EVP recommender model. The composite system-level EVP recommender system was used to predict the best time to activate EVP at any given intersection within the ERV route, in real-time as the ERV traverses across the corridor.
All steps were initially developed in the RBC only environment with final testing in the SILS environment.
The three main steps of the ML development process presented in this section are: 54

(1) Finding the optimum preempt solution,
(2) Developing the ML approach to be used, and
(3) Training the ML using the optimum solution.
In the first step, the optimal EVP decisions for each intersection was determined, accounting for network interactions and varying ERV entry times, using a binary search method involving multiple simulation runs. The aim of Step 2 was to build an intersection level supervised learning model for preemption determination that relies on traffic features (e.g., distance of ERV from given intersection, current signal state, etc.). Traffic features are chosen such that they may be realistically obtained in real-time, given the existing state of CV penetration. The ML algorithm at each intersection was executed each time step (a one-second here) determining whether to implement a preempt call at that intersection. Several supervised learning algorithms were tested. The algorithm, along with a combination of its hyper-parameters, that best fit at an intersection (i.e., best travel-time - preemption duration trade-off) was chosen for that intersection. In Step 3 the optimal solutions from Step 1 are used to study the variation in solution patterns as a function of the features.
While this procedure provides the "optimal" EVP solution for a system of intersections, it was found that when approaching the optimal solution, a very small improvement in ERV travel time could require significant increases in the amount of preemption, and thus much higher side side-street delays. It was observed, in certain instances, that to achieve
55

a gain on the order of 0.1 seconds in ERV response time through an intersection, preempt could increase on the order of 100 seconds. To avoid such instances, a more relaxed optimization criterion is placed where the search is limited to an ERV intersection-level travel time within 2-seconds of the lowest achievable travel time.
The EVP problem was solved at a system level (i.e., for a designated ERV route) as a combination of the solutions at the individual intersection level. An EVP model was created for each individual intersection, and once the ERV indicates the route it wants to follow, the intersections comprising that route trigger their individual intersection level EVP recommender models concurrently to generate the EVP trigger times at each intersection. However, a drawback of this approach (i.e., ERV route-level) is that any failure at the sub-system level (such as gridlock) would not be incorporated in the model. To account for such a scenario, and to ensure safe passage of ERV, an EVP call is made at the predetermined "minimum failsafe" distance of the ERV from the signal, in the case that the ML model does not make the EVP call before the ERV reaches that failsafe distance. Hence, for a given scenario, the training matrix is only created either until the ERV reaches that minimum failsafe distance or until an EVP call is indicated per the "relaxed" optimal solution recommendation.
Training and Validation of the ML Model
Before training the ML-model, the available simulation data was split into training, validation, and testing data based on a 60%-20%-20% basis. The rationale and approach for this data split are discussed in Appendix B. Since the ML model was to be trained to
56

identify the timing of a discrete event (i.e., the proper time to make the preemption call to each signal) it was necessary to develop a "soft-labeling" approach that developed an appropriate "penalty function" that weighted the results based on the overall effect on the final outcome. The development and application of this approach is also discussed in Appendix B.
Choice of the Best model
The problem to be solved consisted of two simultaneous sub-problems: 1) the need to identify situations where no EVP is needed and, 2). if EVP is needed, when the call should to be made. Hence, there are three parameters to control:
(1) No Preempt Threshold which helps the model to learn cases EVP is not required, (2) The form of the evaluation curve from 0 where EVP is not needed up to 1 where an EVP
call is required, (3) The cutoff: The value that determines if an EVP will be called along the evaluation curve.
When the EVPNS goes beyond the cutoff, an EVP call will be made. (a)
Training the model
Multiple combinations of these parameters were used to make the necessary modifications to the training and validation matrices. Supervised-learning was performed to solve this regression problem. Multiple Machine Learning Algorithms were chosen
57

and tested, namely, (1) Support Vector Regressor (SVR), (2) Neural Network and (3) Random Forest Regressor. Each approach has its own merits and limitations and the testing of multiple methods ensured that the overall ML model would not be compromised by in improper selection of learning algorithm. The technical aspect of each of these methods is discussed in Appendix B.
ML Model Evalution
For each intersection, multiple trained models were chosen, combing soft-labeling curve (linear, convex exponential, concave exponential, quadratic), ML-algorithm, and other parameter values. These models were trained on the training data with the process described in Appendix B and then were tested on the validation dataset.
Choosing a model as the `best' model for a given intersection requires a definition for the errors that we are trying to minimize. For this task we considered two conditions: (1) the duration for which EVP call is made, but was NOT needed (i.e., a False Positive) and (2) duration for which EVP call was needed, but the model did NOT activate EVP (i.e., a False Negative). The first determines how much extra EVP is given by the model in question that, in turn, will increase delay on the conflicting approaches, the reduction of which is a secondary objective of the project. The second indicates a missed/delayed EVP call, which can cause additional delay in ERV response, the reduction of which is the primary objective of the model.
58

Hence, to be consistent with the primary objective of the study, false negatives are weighed ten times more than the false positives. This is a decision made based on how much priority a strategy seeks to give to having excess preempt time at the cost of delay to the side-street traffic. This factor could vary based on the general demand in the conflicting movement, time of the day, and general history of what routes are more in demand in emergency conditions. An additional factor in this experiment that plays a role is a True Positive (i.e., an EVP call was needed and was made correctly). The mixed weighted number created with the False Negatives and the False Positives is normalized by the number of true positives. Hence, the best model is defined by a variably weighted factor defined as follows:

. - + -

- - =

-

()

The optimization goal is to have the lowest score possible. In simpler terms, the factor is expected to let one know "how much does a given ML-model get wrong out of the number of seconds the model gets the EVP call right". With that score in place, the top few model-combinations giving the lowest EVP-model-score were chosen and used to run the scenarios in the validation data, as another level of the selection process.

The three models chosen based on their low EVP-model-score were used for further evaluations had several common features and that there was negligible ERV intersection travel times between models (Appendix A). Despite these small differences, the model with the lowest ERV travel time was selected for evaluation. The parameters for the best EVP recommender model for each intersection are provided in Appendix B along with a
59

discussion of the differences between models. For all intersections considered, the Neural Net Regressor model was identified as best.
Results
In the previous sections, the ML strategy has been discussed at the individual intersection level. The following section will focus on the integration of these individual component ML models into an overall ML system model and the testing of this model at the network level. As for the DP algorithm, the individual ML-models were run on the seven intersections concurrently with the ERV placing an EVP call to the PIB @ Medlock Bridge intersection upon entry into the network with the seven remaining intersections along the ERV route having their EVP called based on the ML algorithm.
As before, a 60-20-20 train-validation-test split was performed with 32 of the 160 scenarios set aside for the test-case with the remaining used for the training and validation activities. Model sensitivity experiments were performed on these test data and focused on: (1) Changes in driving factors of the ML-model, viz. random seed, ERV arrival time, etc., and (2) Adjustments in the experimental strategy, such as multiple ERVs in close succession, testing it with SILS architecture, etc.
Model Performance The overall network-level performance of the DP and ML algorithms are compared with the "relaxed optimal" (i.e., perfect knowledge) solution in Figure 5.
60

Figure 5. Graph. Comparative Results in Violin-plots: ML-Prediction model vs other EVP call experiments on the test-data based on ERV Travel Time (s)
While the ML-model (yellow) does not provide the optimal travel time as compared to the absolute global optimal solution (represented by the dark green violin plot), it does provide comparable overall travel time compared to the ideal solution (in grey), which is a `relaxed' optimal solution. Based on a KW-test, it was concluded that that the ERV travel-time in the `relaxed' optimal solution (in grey) and the ML-model driven solution (in yellow) are not statistically different.
Another aspect that needs to be considered is the length of the preempt call placed in each strategy. A longer EVP call can produce good performance for the ERV but be suboptimal for the conflicting routes. Figure 6 shows a violin-plot representation of the total preempt duration for the seven intersections subjected to the ML-model (yellow), when compared to the `relaxed' optimal solution (grey), `extreme' optimal solution (dark
61

green) and the DP solution (pink). Applying the KW-test here, it is found that the preempt duration with the ML-model and the one given by the `relaxed' optimal solution, are not statistically different. This provides initial evidence that a trained ML-model, working with limited data input, can perform as well as more data-intensive strategies.
Figure 6. Graph. Aggregate Preempt Duration on seven interchanges at PIB for Different Experimental Strategies
The mainline and side-street travel-times for non-ERV vehicles for the different strategies are shown in Figure B16. Figure B16(b) indicates that the DP solution has more side-street travel disruption compared to the ML-EVP model with the ML-EVP model (yellow) travel-time for non-ERVs adheres closely to the corresponding traveltime for ERVs from the "ideal" solution for mean and median travel times. This difference is statistically significant based on a KW-test. This provides another strong argument for using the EVP-ML-model in lieu of heuristic DP model. Figure B16(a) also indicates an improvement in travel-time for the non-ERVs that share the same trajectory
62

as an ERV however these differences are not statistically different between the DP and ML models.

(a)

(b)

Figure 7. Graphs. Non-ERV Travel time Variation with respect to ERV trajectory or Preemption Event: (a) Along main-line PIB, (b) All side-street vehicles since an EVP event.

Model Sensitivity Analysis

The ML-model was built using a microscopic simulation model that was calibrated to reflect certain traffic flow and ERV behaviors that were observed in the field. While this process gives a certain degree of validity to the results produced by the simulation model, there are assumptions regarding the behavior patterns of vehicles in the simulation model, which might be violated in the real world under certain circumstances. For example, the simulated ERV never crosses the centerline, or goes over the shoulder, while real world ERVs do perform such maneuvers.
63

Another issue that can arise is overfitting (use of too many variables) which makes the trained model too specific to the training conditions. Sensitivity analysis can aid in determining the presence of overfitting by analyzing the sensitivity of the model outputs to changes in the underlying variables. Several aspects of the model can be adjusted for this type of analysis with the choice of the parameters guided by the variables that played a significant role in model training. The specific model sensitivity tests that were undertaken were:
(a) Varying the Random seed: The random seed value in a VISSIM simulation initializes the random number generator controlling the stochastic variations in various traffic parameters, such as vehicle arrival, speed, acceleration, travel volume, etc. Two simulation runs with identical input files will show difference in those traffic characteristics when run with different random seeds. Hence, varying the random seeds would test the response of EVP ML-model to slight stochastic variations in the traffic conditions that could partially replicate the variations that are observed on different days.
(b) Scaling down Traffic Volume: The simulation model used for testing was based on the PM peak hour. Since the models were tested (and trained) in a high congestion scenario, there is a possibility of obtaining different results if the network volumes were lower. Hence, this condition was tested by running the simulation with the network wide volumes scaled down by 0.5X and 0.75X in two separate experiments, keeping all other elements identical (e.g., no changes in O-D patterns or signal timing).
64

(c) Varying ERV Entry Time into the PIB Network: One of the primary inputs in the simulation model is the time that the ERV is introduced into the network. In the baseline model the ERV was introduced at approximately the one-hour mark of the simulation (i.e., approximately 5:30 PM). To test the sensitivity of models to this factor in the resulting models, the time the ERV was introduced into the network by 800 seconds (i.e., five signal cycle lengths) earlier and later than for the baseline runs keeping everything else in the modeling constant.
(d) Introducing two ERV Vehicles in Tandem to the same Destination: Situations can arise where more than one ERV approaches an intersection. They could be arriving on the same approach or on conflicting approaches at the intersection. This experiment focuses on the case where two ERVs are approaching the intersection from the same direction and tests the efficiency of the model under the concurrent EVP call scenario. Figure B20 shows trajectories of ERVs in tandem with two different levels of gaps between them of 10 and 800 seconds.
(e) Varying Entry Point of ERV: One of the factors in the input matrix for ML-model training is the cumulative distance covered by the ERV in the network. This variable was designed to establish how far away the ERV is from a given interchange based on a consistent entry point. To examine the generalizability of this parameter, the original route was split into two partial routes for ERVs to create two separate experiments: Path 1: Medlock Bridge Road to Pickneyville Park, Experiment 2: South Berkeley Lake Road to Howell Ferry Road.
65

A more detailed discussion of each of these sensitivity tests along with results from each, are provided in Appendix B. Only one of the tests produced significant results, that of the two ERV arriving in tandem.
Figure 8 shows how the ERVs in the tandem scenario affect the average speed of the leading and following vehicles in 16 different cases, with gaps ranging between 10 seconds to 960 seconds between the ERVs. As general intuition would dictate, the leading ERV will not be affected by the current strategy and the model's effectiveness will primarily be tested by how the EVP calls clear out traffic for the following ERV.
ERVs with gaps between them of 90 seconds or less show no meaningful differences between the speed profiles of the leading and following ERVs as both are served with a single `long' EVP call at all/most intersections. Focusing on the gaps highlighted by the orange box, the speed of the following ERV begins to be reduced reaching a minimum with about a 200s gap before beginning to recover. The most likely explanation is that the following ERV only makes its EVP call after the call of the leading vehicle had been completed thereby introducing delay associated with the signal timing transition that occurs between two EVP calls. For example, after completion of the first call the signal controller may need to wait for the service of the minimum green of the current call. At longer time intervals, the effect diminishes until the two ERVs calls can be viewed as independent.
66

Figure 8. Graph. Speed variation of Leading and Following Vehicles for ERVs in Tandem Experiment with varying gap between the two ERVs.
Model Testing in a SILS Environment Since the EVP ML-models were trained using VISSIM's RBC as the signal controllers, the models were also tested using software emulations of actual field controllers used at intersections throughout the State of Georgia. This SILS architecture was built using VISSIM and MaxTime and is the same configuration used to test the DP method. The MaxTime signal controller software also features a different exit strategy referred to as the "Queue Recovery" exit-transition. "Queue Recovery" keeps track of the EVP calls and recalls in the detector and serves the longest waiting phases. This feature was designed to offer an effective way of handling conflicting movements in the aftermath of an EVP call. The performance of the "Queue Recovery" transition was included in the tests along with the default exit-transition strategy.
67

Figure 9 presents the ML-model results with SILS as well as the corresponding results with the internally programmed VISSIM RBCs. On average, the predicted speeds with the SILS, while similar to the results with VISSIM RBCs, were consistently 1-2 mph lower for all the tested strategies when compared to their RBC counterparts. However, as shown in Figure 9, the predicted improvement for ERV travel times for the ML-model over "No Preempt" cases was approximately 115 seconds for both the SILS and RBC architectures. Hence, it appears that the ML-model strategy is not affected any differently than the other EVP strategies, or no-preemption case, by changing from the RBC architecture to the SILS architecture.
Figure 9. Graph. Comparison of ERV KPIs under SILS architecture compared to RBC model based pm ERV travel times.
Another aspect that was investigated was the effect of using Queue Recovery exittransition. A Queue Recovery exit-transition prioritizes the movement that has more queues resulting from the EVP call and serves the corresponding phase. Figure 10 shows
68

how different EVP methods cause varying travel times on the side-streets when EB and WB movements at PIB are aggregated for all seven EVP-programmed intersections going from South Old Peachtree Road to Howell Ferry Road. The figure shows the distributions of side-street travel times and the progression of those with passage of time. As we see, the DP solution and ML-Predicted solution (blue and orange respectively) non-ERV travel-times under default exit-transition settings are consistently higher than the Normal exit RBC setting (light green violin-plot). The effect is reduced as time exceeds two signal cycle lengths (2*160s). With the Queue-recovery setting of SILS (cyan violin-plot), the side-street non-ERV travel times are significantly lower and is somewhat lower than the RBC driven scenario. In contrast ERV speeds do not receive any meaningful adverse effect from the Queuerecovery exit transition (green violin plot: average speed: 39.4 mph) when compared to the default setting SILS case (turquoise violin-plot: average speed: 39.7 mph).
69

Figure 10. Graph. Comparison of non-ERV Side-street travel-time after an EVP Call: Variations of EVP call strategy under SILS Architecture
While the difference is negligible under the current strategy, the effect on any difference from this remains to be studied if actual vehicle pull-over behavior in response of ERV is simulated within the model.
ML-model Runtime discussion The experiments in the previous section showed that the ML-model functions in a robust fashion under various scenarios and shows promise for transferability of the methodology into other networks and different traffic flow conditions. It is important, however, to also ensure that the model can be incorporated in real-time field deployment setups. Hence it is important to test the computational efficiency of the model and ensure the feasibility of running the model faster than real-time so that the decisions can be obtained in time to
70

act. These tests, discussed in Appendix B, show that a relatively modest workstation is capable of running both the RBC and SILS implementation of the ML model faster than real time.
INCORPORATION OF DRIVER PULLOVER BEHAVIOR INTO VISSIM
Overview
The first two tasks studied the effects of EVP on ERV travel times through microsimulation of an ERV moving through an existing network and primarily focused on the ERV's actions and route choices without consideration of any accommodation actions that the non-ERV might undertake. This task focused on ways to improve the accuracy of the model by incorporating the microscopic interactions that occur between ERVs and non-ERVs in the real world due to the mandated requirement of non-ERVs to shift lanes to the right or pull-over to make way for an ERV. Incorporation of this behavior into VISSIM requires modification of the default driver behavior in VISSIM to reflect typical observed behavior of vehicles during encounters with ERVs. This process is described in detail in Appendix C, and is summarized here.
The study focuses on two primary aspects of this ERV/non-ERV interaction. Unlike a normal vehicle, in congested conditions an ERV is expected to try to move over to the left-most lane to try to pass slower vehicles. The non-ERVs are mandated to move over to the right lanes and pull over safely onto the shoulder when possible, to allow the emergency vehicle to use the left lane or left shoulder to overtake the non-ERVs. The
71

lane-shifting and "pull over to the shoulder if needed" behaviors are jointly referred to as "pull-over behavior" and incorporation of this behavior into VISSIM was the primary focus of this portion of the study.
There are several that can affect how drivers may react to the presence of an ERV on the road. Drivers can be expected to vary significantly in both their perceptions and reaction times to an ERV. Moreover, driver compliance with pull-over can vary as well. People may be unaware of the ERV until the vehicle comes within several feet, or they could be reluctant to give up their positions on the road for fear of being taken advantage of and overtaken by other non-emergency vehicles on the road.
Previous studies have acknowledged the importance of modeling these interactions and a few of them have studied the operational and safety impacts. In a recent study, Corts and Stefoni (2023) examined and simulated real-world behavior from video collected in Chile. The upstream and downstream distance ranges within which the ERVs are expected to influence a lane change for the non-ERVs were estimated from manual analysis of the video and was used to model the behavior of non-ERVs in the simulation model. It is however important to note that the number of data points on which the estimates were derived was quite small and limited in scope.
Implementations of ERV-traffic interactions have been developed on several simulation platforms. Zhang et al. (2009) modeled driving behaviors of non-ERVs alongside the ERV movement logic within CORSIMTM. The open-source SUMO platform was used by Weinert et al. (2019). Corts and Stefoni (2023) used PARAMICS.
72

This effort used VISSIM and its External Driver Module (EDM) component to develop a realistic model of ERV/Driver interactions and examines the effects of preemption on ERV and general traffic travel times.
Previous Studies of Driver Behavior
An extensive study conducted by Weinert and Dring (2015) surveyed 252 drivers of fire engines, ambulances, and police cars in Germany to better understand the risks involved during an emergency response. The survey provided strong evidence that non-ERV drivers have varying levels of reaction and response times to the presence of ERVs. Additional video data complemented their efforts to model a rescue lane approach in response to the presence of an ERV.
Buchenscheit et al. (2009) also followed a similar methodology, in gauging the benefits of a proposed emergency vehicle warning system. However, the majority of these studies have investigated the impacts in the context of a V2X or a V2V communications (Buchenscheit et al., 2009; Lidestam et al., 2020; Savolainen et al., 2010). In comparison to a traditional system without communications between vehicles, the reactions of drivers in any system with V2X capabilities tend to be quicker.
For this study task, the focus was on simulating driver behavior without any assumptions about the existence of connected vehicle infrastructure. For the results presented here, the models assume that the ERVs have the capability to invoke a preemption call either with V2I communication or with regular cellular communication between the ERV and the
73

signal infrastructure. No other forms of communication between the vehicles was assumed to exist within the simulation.
Observed Driver Behavior in Georgia
To capture behavioral aspects of drivers along the corridor, video data would have been ideal. However, video data was not readily available for use in this project. As an alternative, the research team sought input from people directly involved in and responsible for the operation of ERVs along the test corridor, including the county's Fire Chief. The responses from this focus group laid the foundations for the modeling assumptions. The primary questions posed to the focus group are listed below, and the conclusions from the responses are summarized in the next sub-section.
Questionnaire When their sirens are on, do fire trucks always use the left lane or left shoulder? Does the emergency vehicle ever cross across the center of two lanes? Under
what conditions? How often (approximately) does this happen? Does the right shoulder ever need to be used? How often do you come across drivers who are unclear about what to do next? Is there a significant rate of pullover non-compliance? How about compliance when the ERV light is red at intersections? How far in advance do drivers begin to shift lanes?
74

In a similar vein, how long do drivers wait to return to their lane once the EV has left?
Does the presence of an EV cause the pace of other traffic to slow down? Do drivers typically leave just enough room for EVs to get through or do they
often make complete lane changes? What specific procedures do EV drivers take when they approach an intersection
with oncoming traffic on cross streets? Do the drivers on the cross streets always slow down to make room for it? Does pull over compliance differ significantly on single-lane roads versus multilane roads? What actions are recommended for other drivers to take while an EV is merging with their link?
Based on the results from the focus group, the following general conclusions as to ERV and non-ERV vehicle interactions along the test corridor were established:
The ERV operators have a defined set of protocols which they must adhere to during emergency response. Despite this, most responses are very incident specific, and may call for non-standard procedures. An example of this is the use of non-paved areas on the road for traversal. Although such a maneuver is not recommended, it can be done if the circumstances deem it necessary.
The overall level of compliance for pullovers is close to 100%. Most vehicles do follow the law and move a lane over, but the time taken to respond can vary
75

greatly. There may be numerous reasons for such variability in response time including distraction and low levels of alertness, lack of available space for the drivers to move, etc. The ERV drivers have been trained to be patient and give the drivers time to execute their response. There are situations where vehicles make a lane change to the left instead of the right, or pull over to the left shoulder, but these are usually observed on multi lane roads, near the intersections. The pull-over responses can usually be seen up to 5-10 car lengths (100 to 200 ft) downstream. The responses of the ERV and the other vehicles are heavily influenced by road geometry, such as the presence of a median, the presence of a left shoulder, etc. Lane partitioning or a "rescue lane" formation is usually observed in congested scenarios, where the ERV has no choice but to go through the middle of the traffic. However, ERVs try and use the leftmost lane as much as possible on their route to the emergency. The general mandate for road users in Georgia is to yield the Right of Way by safely and quickly moving to the right-hand side of the road and come to a complete stop. If possible, they are to pull over to the right shoulder or the nearest available space to allow the emergency vehicle to pass.
These conclusions were used in the modeling assumptions in subsequent modeling efforts.
76

Modeling Components
COM The component object model (COM) interface of VISSIM allows users to interact with the model via scripts. Any required information can be accessed and relayed during the runtime of a model. The scripts can be written in several languages and platforms (e.g., JavaTM, MATLAB, PythonTM etc.). For this study, the scripts were developed using PythonTM.
EDM External Driver Models (EDMs) are additional components that are used to extend the functionalities of VISSIM. EDMs are designed to enhance the current functionalities of VISSIM and give the user more control in "fine tuning" microscopic behavior. Two independent EDMs were designed to handle both behavior classes. These EDMs overwrite the existing defaults within VISSIM. The models are written in C++ and interfaced with VISSIM via Dynamically Linked Libraries (DLL). Throughout a simulation run, there is a constant exchange of data between VISSIM and the DLLs. The DLLs receive the vehicle data and return necessary instructions if defined conditions are met. The use of DLLs allows extensive control over car following models and lane change procedures. The general linkage and data exchange process between VISSIM and the EDM are visualized in Figure 11 below.
77

Figure 11. Matrix. Relationship between the EDM and VISSIM
Model Parameter Adjustments
Several modifications and associated assumptions were required for modelling the ERV and non-ERV behavior to assimilate the results of the focus group study. The details are provided in the following subsections.
1) Vehicles impacted: Vehicles in VISSIM, by default, "see" two vehicles in front of and behind, and two lanes to either side (i.e., expected behaviors will not be visible instantly further downstream or upstream). This area of impact can be expanded by defining new indices and variables using the User Defined Attributes feature available in the recent releases of VISSIM and referencing these values in EDM.
78

2) Impact Distances The primary trigger for initiating pullover behavior is the proximity of vehicles to the ERV. At every timestep, the vehicles check behind them to spot the presence of an ERV. If an ERV is detected, and the distance conditions are met (within 150 ft ), the vehicle is directed to perform a lane change to the right. Similarly, the action of rejoining the lane will only be executed once a minimum headway distance of (330 ft) is satisfied.
3) Speeds ERV models have a custom desired speed distribution, and generally drive no more than 10 mph over the speed limit. Their desired speed is user dependent and can also violate speed limits if required.
4) Separate EDMs Due to the clear separation of expected behaviors on the road, a decision was made to model general vehicles and the ERVs separately. This was done via two separate vehicletype specific EDMs which would function in tandem at every timestep.
Model Design:
The general modeling approach for the pull over behavior model is described below:
1) User Defined Variables are created for the surrounding vehicles, to overcome the limit on the number of surrounding vehicles visible to any given vehicle: a. Category/Type 79

b. ID c. Distances d. Speeds 2) Each vehicle constantly checks their surroundings for the presence of an emergency vehicle. Distance constraints are also applied. 3) If the conditions are met, further actions are controlled by the EDM. 4) Appropriate responses are applied via EDM: a. Lane changes if necessary b. Speed reduction 5) Once the EV has passed, minimum distance and presence checks are completed before passing the control back to VISSIM.
Pullover Frequency
The expectation that all road users will be uniform in their response to an ERV is unrealistic. Thus, each interaction between an individual vehicle and the ERV is modeled separately with unique responses in each case. These choices and responses result of several factors which vary from individual-to-individual, and from situation to situation. For example, reluctance to yield or a lack of alertness can cause delayed responses or drivers may produce unusual responses including abrupt lane changes. These human factors complicate the process of modeling pullover behavior as do environmental and surrounding factors such as traffic density, availability of space, position of vehicle in queue etc.
80

The pullover model developed in this task, uses stochastic variability to simulate this variability. Variability was to be a function of the speed of the vehicle, which in turn loosely relates it to the current traffic density in the corridor with drivers more likely to delay their pullovers under denser traffic conditions. With that assumption, a focus was placed on the frequency of pull-over for vehicles traveling under 5 mph. For vehicles traveling faster than 5 mph, it was assumed they would move as soon as possible. It was also assumed that every vehicle did eventually make a lane change for the ERV (zero non-compliance in line with the focus group results). This behavior was implement using an exponential distribution curve with a mean at 2.5 seconds to model the variation of pull-over delays in the vehicles and generate and assign delay values for each vehicle that meets the pull-over condition and was traveling under 5 mph. A maximum value of 15 seconds was also used as a cutoff. More details on this portion of the model is presented in Appendix C. The next section describes the comparison of results from the VISSIM simulations with and without this custom pullover behavior.
IMPACT OF SIMULATED DRIVER PULLOVER BEHAVIOR ON EVP STRATEGIES
No Pullover versus Pullover models
The primary focus of the study was to reevaluate the impacts of preemption strategies by introducing realistic non-ERV driver behavior in the form of pull-over behavior to
81

improve model accuracy. The effects of the pullover behavior are examined under both conditions, with and without preemption for an emergency vehicle in the corridor. Figure 12 shows results of incorporating the vehicle pullover model into the previous VISSIM cases. The pullover results predict significantly lower travel times for all scenarios with a reduction of approximately 80 seconds between the two no-preempt cases, and about 40 seconds in the preemption scenario. Clearly the impact of the pullover is visibly more significant when there is no preemption. These results also provide an adjusted effect of preemption. The decrease in travel times of ERVs estimated due to the introduction of preemption was on an average 140 seconds without pullover and about 90 seconds with pullover included.
Figure 12. Graph. A comparison of baseline cases without (left) and with (right) pullover behavior using ERV travel times
82

Figure 13 shows the boxplots for the travel times of non-ERV vehicles in the system binned over 160 seconds intervals (1 cycle length) starting from the entry time of the ERV in the corridor. It is clearly seen that with the pull-over behavior there is a significantly higher disruption (larger delays) for the non-ERVs on the mainline in the no-preemption case. However, with preemption the disruption on the non-ERVs is much lower. However, in both cases, with and without pullover, the impact of preemption on the non-ERVs tends to disappear by around the 7th signal cycle from the entry of the ERV.
Figure 13. Graph. Non-ERV travel times observed since the entry of the ERV in the system without pull-over (Left) and with pull-over (Right)
83

Non-uniformity in Pullover To investigate the impact of the assumption of delay exhibited by drivers in pulling over in response to an ERV, a scenario was modeled with an idealized pull-over behavior where vehicles responded immediately to the ERV by pulling over. Figure 14 shows a comparison of the travel times of the ERV for the case without pullover and with idealized pull-over.
Figure 14. Graph. The ideal/non-realistic pullover behavior compared to the non-pullover models
The ERV travel times without preemption in the pull-over enabled scenario show a reduction of almost 170 seconds, which is better than any of the preemption cases in the
84

pull-over disable scenario. Essentially the idealized pullover creates a path of noresistance for the ERVs and would theoretically eliminate almost all delays experienced by an ERV, except for the time lost by the ERV while coming to a stop before crossing a red signal. Further research needs to be done where field data for trajectories of ERVs are used to fine tune the delay distribution used to replicate real world variations in pull-over delay.
ANALYSIS OF OPTIMAL TSP STRAEGIES
Background
In coordinated arterial systems when buses stop to serve passengers, they typically fall out of the progression bandwidth and will likely not traverse the intersection during the current green split. Transit Signal Priority (TSP) aims to provide transit vehicles a free flow path through the intersection, or at least reduce wait time. A summary of this previous work is provided in Appendix D.
TSP system design involves balancing tradeoffs between providing green time to the priority movement, maintaining arterial coordination, and minimizing the impact to conflicting vehicle level of service. TSP is generally implemented at a corridor level, where a transit vehicle will request priority at a given downstream intersection. The priority call is received by the downstream intersection signal controller which, depending on the projected arrival point in cycle, determines if there is a need to implement a TSP signal timing change and what TSP strategy to implement. TSP
85

performance is affected by a wide range of parameters and conditions, including congestion levels, bus headways, bus stop location, detector location, green extension limit, bus arrival time within the cycle, bus stop dwell time, and TSP strategy selected. There are wide variations in the reported benefits and disbenefits of TSP. A review of the literature (Appendix D) also shows that uncertainty in the arrival time prediction, that is, the Estimated Time of Arrival (ETA), from the point of priority request to the stop line remains a challenge in TSP implementation and effectiveness. Arrival times can vary widely depending on traffic conditions (congestion) and bus stop dwell time. Several studies report reduced TSP benefits/effectiveness as congestion increases and have been reported to be more effective for far-side bus stops (bus stop is immediately downstream an intersection) compared to nearside stops (bus stop is immediately upstream an intersection).
According to a survey by National Academies of Sciences Engineering and Medicine (2020), current TSP field implementations have a mix of fixed location and Automated Vehicle Location (AVL) detection systems. Several recent research studies, as well as a few pilot tests. In these studies (Appendix D), CV data is used in ETA predictions as well as to enable improved evaluations of TSP performance. Despite the significant progress made by these studies, there is still significant effort required to reach field ready implementable TSP solutions using CV technology.
86

Problem Definition and Objectives
This task used a VISSIM simulation environment, although a different model (discussed later) than that in Tasks 1, 2, and 3, to evaluate the performance of TSP strategies and establish the critical factors and conditions that affect TSP performance. Critical items assessed were bus stop location (near side vs far side), stop dwell time, ETA, general traffic demand, and signal timing parameter as well as exploring improved strategies for implementation.
Methodology
The methods used to develop the TSP simulations are discussed in detail in Appendix D and will be summarized here. The study used the VISSIM simulation environment to evaluate the performance of TSP under different demand, transit vehicle and transit route parameters, and signal timings parameters. In this task, the VISSIM RBC is taken to provide a reasonable replicate of field signal controller TSP operations as earlier efforts had shown very small differences between the RBC and SILS results. Future efforts may test critical findings in a Software-in-the-Loop environment.
This study performed simulation on variations of a hypothetical network, with the simplest being a single intersection with a far side bus stop and the more complex being a pair of two coordinated intersections with a nearside bus stop at the TSP equipped intersection. The single intersection is assumed to run in coordination, i.e., as part of a system. The initial single intersection experiments sought to test TSP outside the
87

influence of platooned arrivals, allowing for a study of TSP responses not confounded by the arrival pattern. The two-intersection experiments added the confounding influence of platooned arrivals. Two TSP strategies are implemented: (1) both GE (green extension) and EG (early green) available and (2) only GE.
RBC- TSP Algorithm Figure 15 illustrates the RBC's TSP algorithm in coordination. In the figure, Check-in is the time in the cycle when the TSP call is placed and Y+AR is the yellow plus all red for the given phase.
The algorithm operates by granting green extension (GE) or red truncation/early green (EG) to the priority movement or maintaining the current phase timing. The GE, EG, or no priority timing decision is based on the receipt of a call from the intersection's upstream check-in detector and the estimated time within the cycle that the bus would arrive at the stop line. Key inputs include Estimated Time of Arrival (ETA) and maximum GE. ETA is a user input of the estimated/predicted time for the bus to travel from the check-in detector to the stop line. The maximum allowable GE is limited, such that in the cycle after the TSP service there is sufficient time to serve the specified minimum green for each movement. The user can specify any desired maximum GE value below the algorithm's maximum allowable GE.
Upon check-in, the TSP algorithm determines the projected point in the cycle when the bus will arrive at the intersection. If the bus is projected to arrive at the stop line during the phase green time, no action is taken, i.e., the end of green is not altered. If the bus is
88

projected to arrive after the phase green time end, but within the limits of specified maximum GE, GE is granted until the bus checks-out or maximum GE is reached. If the bus is projected to reach the stop line before the start of the phase green time (i.e., arrive on red but beyond point where GE could be applied), EG is requested. Ideally the algorithm will start truncation at a point such that the bus arrives at the stop line without needing to stop. The EG allowed is limited to the remainder of the preceding phase green after serving the maximum of the phase minimum green or a preset minimum called priority minimum green. This minimum green can be the pedestrian time for through movements, if applicable. The algorithm also allows the user to set phases that may be skipped to reach the priority phase earlier. The user can control the algorithm and balance the tradeoff between bus travel time and general traffic delay by (1) selecting appropriate priority minimum greens, (2) selecting appropriate maximum GE, (3) specifying the phases that can be skipped, if any, and, (4) setting the reservice parameter which specifies the minimum time to grant the next TSP request.
89

Figure 15. Matrix. TSP algorithm within VISSIM(R)
TSP Parameter settings The TSP algorithm requires user inputs for several parameters. Three of the most critical are ETA and maximum GE and EG. In all experiments in this study, maximum GE was limited to 20% of the cycle length based on common practice observed in the literature. The base experiment TSP allowed both GE and EG options. Later experiments test the GE only strategy as this approach was undertaken by a subset of agencies in the literature. Minimum greens were set to allow no truncation of the coordinated movement.
90

For side street movements, maximum truncations of 5 and 10 seconds were tested. On the main street 5 second truncation was set for the two left turn movements in all experiments.
Check-in detector location (TSP trigger location) is limited to the distance that can be traveled by the bus at free flow speed. In the far side bus stop experiments, the check-in detector is located after the upstream intersection far side bus stop which becomes a limiting constraint in cases where maximum allowable GE is greater than free flow travel time from the bus stop. This will limit the number of buses that would potentially request GE.
The VISSIM RBC, similar to many field controllers in use today, allows only a single value input of ETA. This value is consistent across all time-of-day plans and free operation. Thus, for all buses in each simulation run, a single value of ETA is set. While this may not be optimal, numerous field deployed TSP systems with fixed location detectors employ the same approach.
Test Network Experiments were performed on hypothetical networks consisting of one or two intersections, utilizing actuated coordinated signal timing. As shown in Figure 16, the base network model consists of a bus route on the main street (E/W) with the bus running in the coordinated movement. There is a fixed location check-in detector 1000ft upstream of the stop line, a checkout detector immediately after the stop line and a far side bus stop immediately downstream of the intersection. The network has two through lanes in each
91

direction for the major street and one lane in each direction for the minor street (N/S). All left turns on the major and minor streets have exclusive turn lanes and only protected signal phases. For simplicity, right turning movements are omitted for both streets.
Figure 16. Simulation. Base network model in VISSIM, Far Side Bus Stop
For the dwell time experiments, an adjacent upstream intersection was added to the model as shown in Figure 17 to enable simulation of platooned arrivals. With arrival profiles controlled from the upstream, the impacts of bus stops including dwell time variability may be more easily isolated and evaluated.
Figure 17. Simulation. Two intersection network model
92

Simulation Parameters The base model considered actuated-coordinated signal timing with a cycle length of 100 seconds. Details regarding the signal timing plans are presented in Appendix D. Preliminary experiments were run in VISSIM to extract saturation headways such that matching values are used in calculating volumes for the different degrees of saturation. Each simulation scenario was run for 30 hours with a bus entering approximately 20 minutes into each hour (Appendix D). Other simulation parameters, including traffic volumes, bus inputs, etc. are provided in Appendix D.
Experiment design To meet the objectives of the study, three sets of experiments were undertaken. The first evaluated critical parameters and conditions that affect TSP performance including: (a) comparing the GE and EG TSP strategies, (b) testing TSP performance at different levels of traffic demand, and (c) assessing the impact of cycle length on both transit vehicles and general traffic. The second set of experiments studied: (a) travel time variability resulting from congestion on the bus route intersection approach, (b) considerations in setting the ETA, and (c) leveraging the capabilities of AVL and CV data to devise improved TSP triggering mechanisms. The third set of experiments are performed on the two-intersection model and are designed to study dwell time variability on TSP performance for both far side and near side bus stops. Details regarding the specific design of each experiment are provided in Appendix D.
93

Results
TSP MOEs and Critical performance parameters and conditions Measures of Effectiveness: The selected MOEs included (1) proportions of granted GE and EG, and proportion of effective GE (i.e., the bus successfully traversed the intersection), (2) transit bus travel time, (3) general traffic delay, and (4) extent/propagation of the delay after the TSP event. Simulation Results General Traffic Delay: Figure 18 presents the delay for the cross street southbound left turn (SBL) movement after a 10 second truncation of the movement green, where the mainline received a 10 second EG. The intersection v/c ratio is 0.95. It is seen that compared to the no TSP case, delay increases and is only able to recover after approximately eight cycles. This figure represents the worst case, i.e., ten second truncation, and not all buses required the max allowable truncation. Where buses received varying levels of EG and GE on average four (4) cycles after each TSP event was sufficient for recovery.
94

Figure 18. Graph. SBL delay extent after 10 second EG
Comparing Impacts of GE and EG: Figure 19 shows the impact of GE and EG, respectively, on bus travel time for a v/c ratio of 0.95. Each simulation run represents 30 bus arrivals. Comparisons between TSP and no-TSP are made for the same bus. That is, if the third bus in the simulation run receives EG, then the third bus in the no-TSP will be used for comparison. It is seen that GE results in higher levels of improvement over no-TSP than EG. This is intuitively reasonable as GE results in a bus skipping the entire red duration whereas as EG reduces the amount of red by the length of the EG truncation. The impacts on other movements are shown in Appendix D.
95

Figure 19. Graph. Impact of GE and EG on bus travel time at v/c ratio of 0.95
TSP Performance under Different Demand Levels:
Table 1. TSP strategies and Effectiveness under different levels of demand shows the TSP strategies and TSP effectiveness for v/c ratios of 0.85, 0.95 and 1.0. In the first TSP strategy both GE and EG are permitted while in the second only GE is implemented. "No Action" is the number of buses for which no TSP was requested, which means either that the bus could make it through the intersection on the normal green phase or the settings prevented the request. "EG" indicates the number of buses that received Early Green while "GE" indicates the number of buses that received green extension. "GE_Eff" shows the number of buses that received green extension and successfully traversed the intersection on the provided GE. GE_NoEff represents the number of buses that were granted GE but maximum GE was reached before the buses checked out/moved through the intersection.
96

97

Table 1. TSP strategies and Effectiveness under different levels of demand

TSP Strategy v/c

EG

GE GE_Eff GE_NoEff No_Action

0.85

119

62

62

0

119

GE & EG 0.95

120

105

100

5

75

1

129

123

100

23

48

0.85

0

62

62

0

238

GE only

0.95

0

109

104

5

191

1

0

134

111

23

166

From the table the following can be noted:
1. From the No_Action and GE columns, more buses need TSP as the congestion increases. This is expected as at lower congestion levels there is less queuing, higher speeds on the mainline, and additional green for the coordinated mainline movement resulting from side street movements gapping out.
2. Green extension effectiveness decreases with congestion as indicated by number of failures (GE_NoEff). For example, 23 out of 134 and 23 out of 123 green extensions for the v/c ratio of 1.0 are ineffective for the two cases of GE only and EG & GE. As discussed later, GE failures are a result of the bus having higher than the set ETA.

98

Figure 20. Graphs. Impact on bus travel time at three levels of v/c ratio, for (a) GE and EG and (b) GE only, on Bus travel time cross v/c shows the bus travel time under the combined strategies of (a) GE and EG and (b) from implementing GE only. Only the buses that receive TSP are included in the results. A higher bus travel time is seen as the congestion level increases from v/c ratio of 0.85 to v/c ratio of 1.0 in the baseline no TSP case. The variability in travel time also increases as congestion levels increase. For v/c of 0.85 and 0.95, GE reduces the travel time for this subset of buses to almost free flow travel time, while for the v/c ratio of 1.0 several vehicles receiving GE still face queuing and therefore the average travel time is higher than the free flow travel time.

(a)

(b)

Figure 20. Graphs. Impact on bus travel time at three levels of v/c ratio, for (a) GE and EG and (b) GE only, on Bus travel time cross v/c

Side Street Delay: 99

Figure 21presents the cross street SBL and SBT movement delay with GE and EG implemented. To allow for comparison across v/c ratios delay is measured for four cycles, starting with the check-in cycle. As expected, the higher v/c ratios of 0.95 and 1.0 show higher delay increases compared to the v/c of 0.85. With a v/c ratio of 0.85 the additional delay is practically insignificant. Additionally, when considering later cycles (five or more cycles after TSP) additional delay to the side streets remains for higher v/c ratios, with the impact of TSP dissipating more slowly than for lower v/c ratios as will be seen in the next set of figures.

(a)

(b)

Figure 21. Graphs. Cross Street Delay from GE and EG for (a) SBL and (b) SBT

Additional results for side street delay times and the extent of side street delay as a function on V/C ratio are presented in Appendix D.

TSP Effectiveness Impact of Cycle Length on TSP Performance: Table 2shows observed TSP strategies and effectiveness for different cycle lengths. As expected, the number of buses not
100

requiring TSP increases slightly as the cycle length increases. This is due to increased mainline green. Furthermore, as the cycle length increases the number of buses proceeding on GE reduces while the number of buses requesting EG is higher. This proportion of GE & EG partly determines the TSP efficiency, with higher GE preferred.

Table 2. TSP strategy effectiveness for different cycle lengths

cycle C110 C130 C150

EG

GE

108

87

125

58

125

45

GE_Eff GE_NoEff No_Action

87

0

105

58

0

117

45

0

130

Bus Travel Time: Figure 22 compares bus travel times with and without TSP for the three cycle lengths. The maximum TSP benefit to buses that receive TSP is provided by the optimal cycle length of 110 seconds. This is likely due to increasing percentages of EG and decreasing percentages of GE as the cycle length increases;

(a)

(b)

101

Figure 22. Graphs. Bus travel time vs cycle length for (a) Only TSP affected buses and (b) All buses
Cross Street Delay:
Figure 23 shows the cross street SBL and SBT delays respectively, for only those bus traversals where TSP was provided. Delay is averaged for 4 cycles after TSP. As expected, in the no TSP case delay increases as the cycle length increases from the optimal cycle length. With TSP, the change in delay after TSP service is lowest at 130 seconds. This reinforces the hypothesis that slack inherent in the higher cycle serves to absorb the increased delay after TSP events. However, as the cycle continues to increase, the general increase in delay due to increased cycle length outweighs the benefits of additional slack to absorb truncation. Thus, TSP is less disruptive to side street traffic at moderately higher cycle lengths than the optimal cycle length. This becomes a tradeoff between accepting higher base line delays (No TSP) and lower delay increases after a TSP event. The decision may further be guided by the frequency of TSP requests. For higher v/c ratios (i.e., 1.0) these trends become move severe, with side street delays during TSP at the lower cycle significantly higher than the higher cycle. This result may lead to a decision to not implement TSP under high v/c conditions.
102

(a)

(b)

Figure 23. Graphs. Cross street delay vs cycle for (a) SBL and (b) SBT
Cross street Delay Extent Figure 24 presents the extent of the change in delay for SBT movement after truncation with a cycle length of 110 seconds. Figure 99 (a) includes the data for any truncation amount up to 10 seconds while (b) only includes the vehicles affected by the maximum truncation of 10 seconds. As discussed earlier the two are very similar as a high proportion of truncations reach maximum truncation for this movement. From the figure, delay increases are experienced up to six to eight cycles after truncation. Similar, although slightly muted, results are seen when the truncation is limited to 5 seconds. For this case the delay increases last for up to four to five cycles.

103

(a)

(b)

Figure 24. Graphs. SBT Delay extent, C= 110s for (a) 0 to 10 second truncation and (b) 10 second truncation only

Similar results for other cycle lengths are presented in Appendix D.

Impact of travel time variability and ETA selection and setting on TSP performance

Bus Travel Time Variability:
Figure 25 shows the variability of travel time for the bus from check in detector to the stop line at a v/c ratio of 1.0. Only buses that check in during green and traverse the intersection on the same green, including GE, are included. Figure 102 (a) is the actual ETA CDF, while (b) plots actual ETA vs time in cycle. The priority phase has a split of 45 seconds that includes 5 seconds clearance time. The phase split starts at 0. The timing has floating offsets and thus the priority movement occasionally receives an early return to green, depending on the side street traffic demand. Data points outside the normal split on the extreme right of the figure indicate bus arrivals at check-in on early return to green.
104

From the data, average actual ETA under these conditions is 30.1, the median is 27.0 seconds, and the 85th percentile is 42.0 seconds. Figure 102 (b) shows the expected trend, i.e., that the buses arriving when the priority phase has just started take longer to reach the stop line as they encounter heavier queuing. Near the end of the phase, several buses approach the intersection at free flow speed, with time estimated between 17 and 20 seconds. Although, even near the end of normal green, there are buses at lower than free flow speed as the queue has not yet cleared, thus setting ETA as free flow time could lead to an ineffective GE.

Figure 25: Bus travel time variability at v/c=1.0.

(a) Actual ETA CDF at v/c=1.0

(b) Actual ETA vs green elapsed at v/c

= 1.0

Figure 25. Graphs. Bus travel time variability at v/c=1.0.

Similar Results for v/c for 0.95 and 0.85 are shown in Appendix D.

105

ETA Sensitivity:

TSP performance results for different ETA selections at v/c = 1.0 are shown in Table 3. From the table it is seen that the number of both GE and EG requests are higher for lower ETA values and decreases progressively as ETA increases. The number of unsuccessful GEs is also higher for lower ETAs and decreases progressively as the ETA increases. At lower ETAs, there are more GE attempts but also more GE failures due to underestimation of the actual bus travel times and consequently the bus failing to pass through the intersection before the expiration of maximum GE. Thus, lower ETAs increase GE opportunities but present a risk of additional unsuccessful GEs, while higher ETAs may have missed GE opportunities.

Table 3. TSP Effectiveness under different ETA values

TSP_strategy ETA

17

20

25

30 GE&EG
35

40

45

50

GEonly

17

EG 138 129 119 115 109 104 96 94 0

GE 121 123 121 112 103 91 80 68 134

GE_Eff GE_NoEff No_Action

97

24

41

101

22

48

99

22

60

94

18

73

87

16

88

75

16

105

69

11

124

59

9

138

110

24

166

106

20

0

134

112

22

166

25

0

126

104

22

174

30

0

114

96

18

186

35

0

104

88

16

196

40

0

92

76

16

208

45

0

92

76

16

208

50

0

69

60

9

231

(a) GE and EG

(b) GE only

Figure 26 (a) shows the bus travel time for a GE and EG strategy implementation, while Figure 26 (b) is GE only. In the figure ETA17 stands for an ETA of 17 seconds and so forth. The error bars represent 95th confidence intervals. All buses are included in the data whether they receive TSP or not, as the requests for TSP are partly determined by the ETA values. There is a trend of increasing bus travel time as ETA values increase. This is mainly because fewer GE opportunities are identified. Additionally, lower ETAs encourage more aggressive truncation which can clear the queue ahead of the bus.

107

(c) GE and EG

(d) GE only

Figure 26. Graphs. Impact of ETA selection on bus travel time (All Buses)

Figure 27 (a) and (b) show the cross street SBL and SBT delay for the different ETA selections with the GE only strategy implemented. The lower ETA values show higher delays due to an increase in granted TSP requests and more failed GE.

(a)

(b)

Figure 27. Graphs. Cross street Delay Vs ETA for (a) SBL and (b) SBT

108

AVL Algorithm Table 4 shows the strategies and effectiveness of TSP implemented with the modified AVL algorithm for the v/c ratio of 1.0. From the GE_NoEff column, comparing AVL and fixed location CI-CO TSP, the failed extensions reduce from 19 to 3 in the GE and EG case and from 19 to 1 in the GE only case. This shows the potential for greatly improving TSP effectiveness using an AVL system incorporating CV data.

Table 4. TSP strategies and Effectiveness with AVL Algorithm

TSP_strategy Det

GE&EG

AVL CI-CO

GEonly

AVL CI-CO

EG 161 153 0 0

GE 105 113 120 131

GE_Eff GE_NoEff No_Action

102

3

24

94

19

24

119

1

170

112

19

159

Figure 28 shows a comparison of the AVL TSP system and the fixed location detector CI-CO for both GE & EG and GE only strategies. Only buses that receive TSP are included in this data. Whereas there is minimal difference between the two systems in the GE and EG implementation, AVL performs better in the GE only implementation. The gain is derived from the reduced green failures. Larger benefits may be expected at higher v/c ratios and for the nearside bus stops where the ETA uncertainty is increased by the dwell time at the bus stop.
109

GE and EG

GE only

Figure 28. Graphs. Bus travel time, Comparing AVL and fixed location check-in.

Impact of dwell time magnitudes and variability Dwell time magnitudes and variability at nearside bus stop: As can be expected the dwell time magnitude at a far side bus stop affects the arrival profile of the bus at the check-in detector. As such, a number of simulations were performed to examine the effect of Dwell time on the effectiveness of TSP. The results from these simulations are presented in Appendix D.

Comparing TSP Effectiveness for Nearside and Farside Bus Stops: Table 5 summarizes the TSP strategies and effectiveness for far and nearside bus stops for the three dwell time distributions (Appendix D). The most informative column is "GE_NoEff," which represents the failed green extensions. The column "% GE_NoEff" represents the percentage of the failed GE compared to the total GE. For the nearside bus stops, GE effectiveness decreases as the magnitudes and variability of dwell time

110

increases. Buses with a higher value of dwell time than used in the estimate of ETA may be granted GE but fail to make it through the intersection.

Table 5. TSP strategies and effectiveness at far and nearside bus stops

stop

TSP

%

No

location Strategy DT EG GE GE_Eff GE_NoEff GE_NoEff Action

DT01 75 129 121

8

GE &

DT02 117 104 97

7

EG

DT03 153 84

81

3

Farside

DT01 0 126 119

7

6.2

96

6.7

79

3.6

63

5.6

174

GE only DT02 0 105 96

9

8.6

195

DT03 0 87

83

4

4.6

213

DT01 61 143 104

39

GE &

DT02 80 146 94

52

EG

DT03 98 156 74

82

Nearside

DT01 0 146 108

38

27.3

96

35.6

74

52.6

46

26.0

154

GE only DT02 0 152 104

48

31.6

148

DT03 0 166 81

85

51.2

134

Figure 29 compares bus travel times for far side and nearside systems. In both cases, travel time is measured for the same distance starting upstream of the far side bus stop location. Thus, travel time includes dwell time at the bus stop in both cases. Additionally, only buses receiving GE are included in the plots. Comparing the TSP and No TSP cases, it is shown that (1) TSP saves the bus more time for far side bus stops compared to near
111

side bus stops, (2) the performance of TSP for the near side is much more variable, and (3) for nearside bus stops, the performance of TSP in saving bus travel time deteriorates as the magnitude and variability of dwell time increases.

(a) Dwell time distribution, DT01

(b) Dwell time distribution, DT02

(c) Dwell time distribution, DT03
Figure 29. Graphs. Bus Travel time for far and nearside bus stops
112

Considering the relative ineffectiveness of TSP at nearside bus stops, green time is wasted at the expense of side street movements at a much higher rate than with far side bus stops. Figure 30 compares of cross street SBL delay for nearside and far side stops and the three dwell time distributions. The plots show cross street vehicle delays in the four cycles after the bus check-in, whether the bus receives TSP or not. It is noted that (1) delay caused by TSP is higher for nearside bus stops compared to far side stops and (2) for nearside bus stops the delay caused by TSP increases with dwell time magnitudes and variability.
Figure 30. Graphs. Cross street SBL Delay, Farside vs Nearside bus sto
113

CHAPTER 4. ANALYSIS AND DISCUSSION
EMERGENCY VEHICLE PREEMPTION WITH DYNAMIC PREEMPTION LOGIC
This task aimed to quantify the impacts of the use of CV technology on EVP by implementing a DP architecture in a microscopic simulation model with ERV travel time as the primary metric. For the cases studied, DP with either exit transition led to an approximate 23% (~140s) reduction in average travel time for the ERV. While not as significant, it was seen that traditional CI-CO also had the potential to improve ERV travel time, with approximately half the benefit seen from the DP algorithm. In the DP scenarios, the non-ERVs sharing the same path as the ERV also received a significant reduction delay as a by-product of the preemption. When considering the side-street there was a preemption-derived disruption in travel leading to increased cross street travel times. However, this effect dissipated quickly.
Since the normal exit transitions were observed to be slightly superior in service to the side streets during the disruption period, DP with normal exit was found to be the most favorable preemption strategy for the network and conditions studied. However, generalization of these results will require additional study of other corridors and under varying traffic conditions with other exit modes beyond those discussed here.
The observed improvements or degradation in travel time are not directly transferrable to other corridors, each corridor with unique characteristics and demands requiring analysis.
114

However, both this and other studies have demonstrated the potential benefits of EVP strategies and the potential for additional benefits with CV implementation.
The work on this task also identified areas for potential improvements in future studies. First, a mix of the main and side-streets along the ERV route versus the mainline-only route used in the study might provide additional insight into network effects. Second, scenarios with arrival of multiple ERVs at an intersection should be explored to better understand the interaction of overlapping preemption request calls. Third, there have been several studies tackling path planning problems for EVP (Gedawy et al., 2008; Saravanan et al., 2017; Zhao et al., 2017). Since it was known a prior that the ERV will only be traveling on the mainline utilizing a designated route, this is not included in the current experimental architecture. Fourth, the comparative analysis of EVP could have been extended beyond just travel time to additional metrics such as average queue-length, cycle failures on the cross streets, etc. to provide more depth to the findings. Lastly, while successful, the DP algorithm developed is a relatively simple heuristic and additional refinement of the algorithm is probably merited.
EMERGENCY VEHICLE PREEMPTION WITH MACHINE LEARNING LOGIC
Task 1 demonstrated how integration of real-time traffic flow information from CV technologies (OBU on the ERV and RSUs at the intersections) could feed ad hoc heuristic "Dynamic Preemption" (DP) logic to improve the movement of the ERV through the intersections. The DP algorithm assumes real-time availability of queue
115

lengths and uses this information to make decisions regarding when to make the preemption call. To bypass the need for this level of high resolution real-time field data an ML approach was investigated, whereby the EVP trigger recommendation will be driven by conventional detection data streams such as vehicle detection from inpavement or roadside count detectors. A supervised ML approach was chosen for the development of an optimal preemption strategy. The EVP recommender models formed for contiguous intersections were combined and run concurrently to create a system-level EVP recommender model. The composite system-level EVP recommender system was used to predict the best time to activate EVP at any given intersection within the ERV route, in real-time as the ERV traverses across the corridor.
The ML model developed showed excellent results, and was able to achieve, on an average, a 42.5 mph speed for the ERVs through the corridor, which is within 10% of the speed-limit. The results confirmed that this performance did not come at the cost of extended preemption durations. The preemption durations for the ML logic were not significantly different from those generated by the DP logic. Similarly, the level of disruption of the non-ERV cross street traffic was comparable.
One of the concerns with the development of such a ML model is the possibility of overfitting the model to a narrow set of criteria. To check for overfitting a series of sensitivity analyses were performed using test scenarios that used variations in random seeds, traffic volumes, and entry time of the ERVs in the network. It was found that the preemption trigger times generated by the ML logic under these varied scenarios
116

provided stable improvements in travel times in the same range as the DP strategy. The sensitivity analyses helped test the robustness of the ML model and establish the feasibility of using ML for developing the logic for generating preemption triggers using conventional vehicle detection data combined with ERV location information and intersection signal phase status data.
IMPACTS OF DRIVER PULLOVER
In this task, the primary focus was on the development of models that would allow for more realistic simulation of preemption scenarios than those currently accommodated by default driving behavior models in microscopic simulation platforms such as VISSIM. Although preemption strategies and their effects on Emergency vehicles have been explored by many, few studies have accounted for the complex interactions between the general traffic and the ERVs. This study developed realistic models for driving behaviors for interactions between ERVs and non-ERVs which replicate the pull-over lane clearing behaviors that are observed in the real world in response to ERVs. External Driver Models were created on VISSIM, and their effects were studied in a real-world corridor. The ERV and non-ERV driver models were based on findings from a focus group consisting of emergency response personnel. In the absence of direct field data to calibrate the pull-over driver behavior model two variations of the model were studied, one with idealized instantaneous pull-over, and the another non-idealized conditions where the pull-over is affected by delays attributable to lack of alertness of non-ERV drivers, indecision in making the required maneuvers to yield to the ERVs, and the non-
117

ERV drivers' reluctance to give up their position in the lane over concerns of being overtaken by other non-ERV drivers leading to delays. The idealized pull-over model essentially created a path of no resistance for the emergency vehicle by cooperatively moving out of the way of the ERV in such a way that the ERV is not required to slow down for the non-ERVs. In this scenario the only delay experienced by the ERV is for its mandate to come to a full stop and checking for cross street traffic before crossing an intersection that has a red signal for the approach used by the ERV. Additional gains in travel time improvement for the ERV tended to be minimal, in the order of 10%, with preemption. To accommodate for realistic response of non-ERV drivers to an ERV, an exponential model was used to represent the distribution of the delays the drivers are expected to exhibit before pulling over. Minimum and maximum cutoff of 2.5 second and 15 seconds were used for the delays. This model showed a more realistic gain in travel times for the ERVs with preemption, in the order of 16%. On the other hand, the non-ERVs on the path of the ERV experienced a disruption due to the pull-over behavior when there was no preemption. However, with preemption, the non-ERVs on the path of the ERV tended to benefit from the preemption, with reduction in travel times for the next few minutes, due to the queue flush initiated, similar to the nopull-over scenario.
118

TRANSIT SIGNAL PRIORITIZATION
Transit Signal Priority (TSP) aims to provide transit vehicles a free flow path through the intersection, or at least reduce wait time. In the literature the effectiveness of TSP is mixed, having been shown to reduce transit vehicle delays, improve reliability and schedule adherence, and mitigate bus bunching in some instances, while having minimal effect in others. TSP systems are typically designed for corridor operations and involve balancing tradeoffs between providing green time to the priority movement, maintaining arterial coordination, and minimizing the impact to conflicting vehicle's level of service. TSP performance is affected by a wide range of parameters and conditions, including congestion levels, bus headways, bus stop location, detector location, green extension limit, bus arrival time within the cycle, bus stop dwell time, TSP strategy selected, and uncertainty in the arrival time prediction, that is, the Estimated Time of Arrival (ETA),
This study uses a VISSIM simulation environment to evaluate the performance of TSP strategies and establish the critical factors and conditions that affect TSP performance. Critical items assessed are TSP strategies, stop dwell time, ETA, general traffic demand, signal timing parameters, bus stop location (near side vs far side), and TSP triggering strategies both with check-in-check-out (CI-CO) detectors and AVL with CV data.
As such, through a series of simulation experiments the following findings were determined:
119

1) Several key findings when considering Green Extension (GE), i.e., extending the current green phase to allow a bus to traverse the intersection, and Early Green (EG) i.e., truncating the red to allow a stopped (or arriving) bus to depart sooner, include: a. GE tends to provide more significant bus travel time improvements than EG. This finding is intuitively rational, as with GE the bus skips the entire red, whereas with EG the bus travel time is only reduced by the amount of red truncation. b. GE tends to have a lesser impact on conflicting movements than EG. The delay for the same number of seconds of extension versus truncation tends to be higher and last for more cycles when truncation (i.e., EG) is granted. This is primarily a result of the ability with GE to return to the coordinated timing plan by proportionally shortening all phases in the next cycle, whereas EG tends to require all phase time reduction to come from only the truncated phases preceding the TSP movement.
2) As congestion increases (modeled by increasing v/c ratios in this study) several findings are noted: a. With increasing congestion, the percentage of buses requesting TSP increases due to slower speeds, increased presence of queuing and side street phases maxing out. b. The likelihood of an ineffective GE increases with congestion, i.e., in an ineffective GE a bus requests TSP and GE is provided; however, maximum GE is reached before the bus traverses the intersection and the bus is stopped. c. When providing both GE and EG the bus travel time is lower at lower v/c ratios. This is primarily a result of increasing percentages of EG as v/c increases, where 120

the benefits of EG is less as previous noted. Also, the percentage of inefficient GEs increases with increasing congestion. d. However, when providing both GE and EG the absolute travel time improvement between TSP and non-TSP is greater at higher v/c's, as the non-TSP travel time increases more rapidly than the TSP travel time as v/c increases. e. Similar trends are seen in the travel time of buses that receive TSP when only GE is utilized; however, the improvements are more significant, further demonstrating the advantage of GE. f. The impact to side street delay increases as v/c increases. This is a result of the reduced flexibility to recover as demand increases for the green time. g. As v/c increases the number of cycles for the sides side street delays to return to non-TSP levels increases. For v/c ratios of 1.0 the side street may not return to non-TSP levels for an extended number of cycles past the TSP event, at least 14 cycles in this study. Thus, the likelihood of the impact of the previous bus still being present when the next bus arrives increases as v/c increases. 3) As cycle length increases (for the same volume set): a. The proportion of buses requesting TSP decreases. b. The proportion of buses requesting GE decreases. c. The effectiveness of GE requests improves. d. Bus travel time averaged across all buses increases slightly. e. Cross street delay due to TSP is lowest when the cycle is moderately higher than optimal (calculated to optimize non-transit) movements. f. Overall cross street delay increases as cycle length increases beyond optimal.
121

g. The number of cycles required to dissipate the effect to the side street reduces as well as the impact per cycle.
4) The following is the impact of travel time variability and ETA selection, where only a single ETA may be entered into the traffic controller: a. A short ETA may result in higher GE ineffectiveness, where the maximum GE is exceeded before the bus traverses the intersection. b. A long ETA may result in GE not being granted in situations where the bus would have been able to traverse the intersection. c. As v/c increases travel time variability increases, making the selection of ETA more challenging. Lower v/c results in a higher percentage of buses at free flow speed and improves the effectives of the lower ETA. d. Overall average bus travel time tends to be lower at the lower ETA. e. Overall impact on the cross street tends to be lower at the higher ETA.
5) When using AVL TSP effectiveness may be improved by mitigating some of the tradeoff that occurs between the high and low ETA values.
6) Dwell time can have a significant impact on TSP, introducing significant uncertainty in the bus arrivals, thus reducing the effectiveness of TSP. Setting of offsets to account for dwell time uncertainty may help improve TSP performance.
7) Near side bus stops introduce more uncertainty in bus arrival time, particularly highlighting the need for an AVL based TSP solution that allows for real-time ETA updates.
122

CHAPTER 5. CONCLUSIONS AND RECOMMENDATIONS
CONCLUSIONS
This first task of this study demonstrates the benefits of using EVP, combined with CV technology, by developing a DP logic and implementing it in a microscopic simulation model. Dynamic Preemption, with either Normal or In-step exit transitions, led to an approximate 20% (~125 s) reduction in average travel time for the ERV when the ERV traversed through a series of eight preemption-enabled signalized intersections on a highvolume congested corridor. The DP algorithm provided a significantly larger reduction in travel time of the ERV than a traditional check-in-check-out detector-based preemption (~55 s) as compared to the no-preemption case, thus providing a strong demonstration of the advantages of a dynamic approach in preemption that leverages real-time location and traffic data.
The non-ERVs sharing the same path as the ERV also received a significant reduction in delay as a by-product of preemption, primarily due to the queue flush in front of the ERV. When the preemption request was on a mainline approach phase, it was observed that there is a disruption in travel behavior on the side streets, leading to increased travel time with preemption. However, that negative impact on average travel time dissipated rapidly over time. There may also be slight advantages and disadvantages to both the Normal and In-step exit strategies in this regard. For instance, the Normal exit strategy tends to serve the non-ERV traffic side streets slightly better when the preempt call is on
123

the mainline, whereas the In-step strategy tends to serve the non-ERV mainline better when the preempt call was on the side-street.
The primary contribution of this effort, however, is in Task 2, that demonstrates the development of an ML-model that is significantly less data hungry than the Dynamic Preemption model and provides superior ERV performance than either DP or CI-CO. The Dynamic Preemption model requires estimates of the queue lengths at the intersections ahead of the ERV. With heavy penetration of CV technology this would not be a challenge. However, while CV technology penetration is ramping up, the available data are still mostly limited to vehicle detection from traditional infrastructure detectors on the pavement. The ML-model based strategy developed here helps bridge the CV ramp up period gap by providing a methodology that works with the traditional data while it leverages some facets of CV technology that are less onerous to deploy, such as limited deployment of CV units at intersections and on board the emergency vehicles.
One of the major challenges of using ML models is meeting the data requirements for training the ML model. The simulation platform was used to develop and demonstrate the feasibility of creating an ML model for EVP operations. Multiple iterations of simulations were used to generate solutions with varying parameters of simulation and ERV arrivals. The ML-model was trained to learn the characteristics from the available set of solutions with the use of factors such as detector data and signal information. The MLmodel produced results that were better than the performance of the DP solution on the same ERV route.
124

To test for robustness and the potential for generalization of the process in other scenarios and traffic networks, a deeper analysis was performed on the effectiveness of the ML-model. Several key parameters driving the experiment, ranging from simulation random seed, ERV arrival time, variations in ERV route, variations in simulation strategy (e.g. SILS, multiple EVs in tandem, etc.), etc. were tweaked to test the sensitivity of the EVP recommender system. The sensitivity tests provided evidence of the robustness of the model and demonstrated the potential for successful application of the model across different scenarios and networks.
Task 3 of this study addressed the issue that the vast majority of previous studies on preemption have failed to account for the complex interactions between the general traffic and the ERVs. The ones that investigated these interactions typically studied the interactions from an idealized perspective to demonstrate the potential benefits of enhanced V2X communication enabled by CV technology. This study developed realistic models for driving behaviors for interactions between ERVs and non-ERVs which replicate the pullover lane clearing behaviors that are observed in the real world in response to ERVs. The study generated External Driver Model dynamic linked libraries that can be integrated by others for preemption studies to integrate pull-over behavior in the models and provide a realistic baseline that will prevent an unrealistically optimistic bias in the results.
Task 4 of this study expanded the effort to TSP. Overall, it is seen that TSP performance is most favorable in lower v/c conditions where far side bus stops are present. The lower v/c and greater distance between the upstream bus stop and the intersection lessens the uncertainty in ETA, which is critical to TSP effectiveness. Green extension
125

also provides the most benefit to individual buses. However, as congestion increases the effectiveness of TSP decreases. On a highly congested corridor, i.e., v/c ratios approaching or exceeding 1.0, it is possible that TSP may become infeasible as the non-TSP movements may have insufficient slack in available capacity to recover from the TSP related green truncation.
Underlying TSP is a balancing of transit and other vehicle demands. The signal timing for optimal transit vehicle performance may not be optimal for other vehicles in the network. For instance, it was seen that slightly higher cycle lengths or adjusted offsets compared to those for demand based optimal signal settings may result in better TSP performance, as well as lower impacts to non-transit vehicles during TSP events. However, during cycles without TSP the delay experienced by the non-transit vehicle will be higher than under the lower cycle scenario. One implication of this finding is that the setting of a corridor's signal timing parameters should reflect the corridor purpose. That is, if the corridor is designated to serve a significant transit function, then base timing parameters should be selected to improve bus travel time. Where transit is not a primary focus then base signal timings such as higher cycle lengths or TSP based offset may not be warranted.
STUDY LIMITATIONS
There are a few limitations of the study that needs to be acknowledged while interpreting the results.
126

The dynamic preemption based logic for preemption makes two significant assumptions. The first is that the path of the ERV is known. With the used of Computer Aided Dispatch systems and the wide availability of navigation software and applications, it is not unrealistic at all. However, the communication of the information between the different systems, such as the dispatch systems of emergency management and public safety departments, and the signal controller software operated by the department of transportation, might not exist in all jurisdictions.
The second assumption for the dynamic preemption logic is about the availability of realtime queue length estimates. While obtaining such estimates is efficient in a scenario with sufficient level of penetration of connected vehicle on-board-units, in reality, a widespread deployment of on-board-units does not exist currently and might take several years to happen. To bridge the gap between now and the time when queue measurements from the field become ubiquitous, the study develops the machine learning model for preemption.
However, the machine learning model also has certain limitations. The current study performed several sensitivity tests to establish the robustness of the ML algorithm under different conditions of variation of traffic on the corridor. However, this does not allow for any conclusions about the robustness of the algorithms when the geometric conditions change. In other words, the study cannot draw any conclusions on whether the algorithm developed using data from one corridor will be usable in another corridor. In fact, given the variables used by the algorithm as input, it is likely that the algorithm will need to be
127

trained separately for each corridor where it is deployed. However, this is not totally unrealistic either, traffic engineers are familiar with the requirement for customization of signal timings for every deployment corridor, especially when any advanced strategy such as coordination or adaptive signal timing is used.
The limitation of the pull-over model development study was somewhat different. This study suffered from the lack of empirical data. The model was developed based on the collective experiences and observations by emergency personnel as captured in a focus group interview. However, for the model to be vetted, it is essential for future research to consider collecting empirical data that can be used to tune the parameters such as the lengths of the impact zones, delay in pulling over, etc., that were currently based on observations that were not collected as part of an organized data collection effort.
A limitation of the task 4 TSP effort is that the study is currently limited to one or two coordinated intersections and a limited number of demand sets. While this study allowed for an exploration of fundamental TSP principles a next step should involve testing on real-world corridor using existing conditions and bus schedules. Such as study should include stepping from a simulated analysis, to software-in-the-loop, to field implementation and evaluation. Finally, it was seen that ETA is a critical parameter in the setting of TSP, particularly in congested conditions. Unfortunately, the controllers considered only allowed for a single ETA parameter. This creates a significant constraint on TSP performance. It was seen that AVL provides a significant opportunity to mitigate
128

this drawback. Future efforts should seek to further expand the AVL algorithm explored and consider test implementations.
RECOMMENDATIONS
Based on the conclusions drawn from this study, the following recommendations are made to assist GDOT decisions, on a project level basis, about EVP and TSP deployments decisions and strategies:
The findings of the study clearly demonstrated the advantages of using a dynamic preemption logic over using a fixed check-in-check-out detector based preemption logic. Using a dynamic logic for preemption is therefore recommended under most circumstances where real-time traffic information is available. In cases where accurate queue estimates are not available from the field, ML models can be used to work with conventional vehicle detection data streams, as demonstrated in task 2 of this study.
The study also compared the impacts of two exit strategies for preemption, normal preemption and in-step preemption. It was found that the Normal exit strategy tends to serve the non-ERV traffic side streets slightly better when the preempt call is on the mainline, whereas the In-step strategy tends to serve the non-ERV mainline better when the preempt call was on the side-street. Hence the choice of the exit strategy will need to take into consideration factors such as the
129

demands and turn ratios at the intersections, the use of coordination on the relevant corridor, etc.
The pull-over study helped validate the results of the EVP study in a more realistic simulation scenario. Even when the non-ERVs pulled over for the ERV, the benefits of using EVP in reducing delays for the ERV was still evident and clearly makes the case for use of EVP at signalized intersections.
Based on the findings from the pull-over modeling study an additional argument can be made for deploying EVP to avoid disruption to the traffic along the path of the ERV. It is clear that pull-over causes disruption to the traffic traveling in the same path as the ERV. The comparison of the non-ERV travel times between EVP and non-EVP scenarios, with pull-over integrated in the model, clearly shows that EVP can minimize the disruption with a relatively minor short term disruption to the cross traffic.
When considering the setting of TSP on a transit designated corridor:
If intersections with a v/c ratio on the order of 0.95 or higher exist, it is recommended to consider slightly longer cycle lengths (on the order of 10 to 30 seconds) to determine if additional slack (unused capacity) in the timing may be obtained. The general traffic delay will need to be checked, comparing the optimal cycle length vs transit cycle length, to determine the acceptability of this option. 130

Where bus stops exist upstream of an intersection offsets that maximize the opportunity for bus passage (include GE time) should be investigated. Intersection dwell time distributions (typically available from the given transit agency) should be included in the modeling of the signal timing.
Where conflicting movement delay is highly sensitive to TSP (typically in higher v/c situations) it may be desirable to limit TSP to GE as GE tends to higher benefits with lower impacts than EG.
While not directly studied in this effort selective TSP may be considered. For instance, only providing TSP to those buses that are running behind schedule.
The selection of low (free flow) vs high (congestion based) ETA should be considered in relation to the corridor objectives. While free flow speed based ETA will likely provide overall better service, where the focus is on congested conditions longer ETA may prove a more suitable option.
Where possible AVL (or other) solutions that allow for flexibility in the selection or application of ETA should be considered.
While beyond the scope of a signal timing only effort consideration should be given, where possible, to the placement of bus stops on the far side rather than near side of an intersection.
131

The above TSP recommendations represent a generalization of the study result, within the stated limitations. As every corridor has unique characteristics, each corridor should be modeled to determine the most effective application of these recommendations. However, key to the application of any transit timing is that transit timing be considered as part of the signal timing objectives rather than an afterthought to be applied by "tweaking" the general traffic "optimal" results.
132

KEY REFERENCES
Anderson, J., Suh, W., Guin, A., Hunter, M., & Rodgers, M. O. (2019). Accounting for composite travel time distributions within a traffic stream in determining Levelof-Service. Journal of Intelligent & Fuzzy Systems, 36(2), 955-965.
Buchenscheit, A., Schaub, F., Kargl, F., & Weber, M. (2009, 28-30 Oct. 2009). A VANET-based emergency vehicle warning system. 2009 IEEE Vehicular Networking Conference (VNC),
Corts, C. E., & Stefoni, B. (2023). Trajectory Simulation of Emergency Vehicles and Interactions with Surrounding Traffic. Journal of advanced transportation, 2023. https://doi.org/10.1155/2023/5995950
Gedawy, H. K., Dias, M., & Harras, K. (2008). Dynamic path planning and traffic light coordination for emergency vehicle routing. Comp. Sci. Dept., Carnegie Mellon Univ., Pittsburgh, Pennsylvania, USA, 189.
Lidestam, B., Thorslund, B., Selander, H., Nsman, D., & Dahlman, J. (2020). In-Car Warnings of Emergency Vehicles Approaching: Effects on Car Drivers' Propensity to Give Way. Frontiers in Sustainable Cities, 2. https://doi.org/10.3389/frsc.2020.00019
Mohammadi, R., Roncoli, C., & Mladenovic, M. N. (2020, 3-5 Nov. 2020). Transit signal priority in a connected vehicle environment: User throughput and schedule delay optimization approach. 2020 Forum on Integrated and Sustainable Transportation Systems (FISTS),
National Academies of Sciences Engineering and Medicine. (2020). Transit Signal Priority: Current State of the Practice. https://nap.nationalacademies.org/catalog/25816/transit-signal-priority-currentstate-of-the-practice
Obenberger, J., & Collura, J. (2001). Transition strategies to exit preemption control: State-of-the-practice assessment. Transportation Research Record, 1748(1), 7279.
OpenStreetMap. OpenStreetMap: Peachtree Industrial Boulevard Corridor. OpenStreetMap. Retrieved January 17 from https://www.openstreetmap.org/directions?engine=fossgis_osrm_car&route=33.9 4488%2C-84.23857%3B34.01961%2C-84.16084#map=13/33.9823/-84.1997 133

PTV, A. (2021a). PTV Vissim- 3D Vehicle Models. PTV AG: Karlsruhe, Germany. Retrieved July 28 from https://company.ptvgroup.com/en-us/ptv-vissim-3dvehicle-models
Python. (2018). Python 3.7.0 Release. Retrieved June 25, 2020 from https://www.python.org/downloads/release/python-370/
Saravanan, R., Paul, P. V., Nivedha, M., & Nirosha, B. (2017). Optimal path planning and traffic signal control system for emergency vehicle--A survey. 2017 International Conference on Computation of Power, Energy Information and Commuincation (ICCPEIC),
Savolainen, P. T., Datta, T. K., Ghosh, I., & Gates, T. J. (2010). Effects of dynamically activated emergency vehicle warning sign on driver behavior at Urban intersections. Transportation Research Record(2149). https://doi.org/10.3141/2149-09
Trucano, T. G., Swiler, L. P., Igusa, T., Oberkampf, W. L., & Pilch, M. (2006). Calibration, validation, and sensitivity analysis: What's what. Reliability Engineering & System Safety, 91(10-11), 1331-1357.
Weinert, F., & Dring, M. (2015, 2015//). Development and Assessment of Cooperative V2X Applications for Emergency Vehicles in an Urban Environment Enabled by Behavioral Models. Modeling Mobility with Open Data, Cham.
Weinert, F., Dring, M., & Bogenberger, K. (2019, 27-30 Oct. 2019). Influence of Emergency Vehicle Preemption on Travelling Time and Traffic Safety in Urban Environments enabled by Innovative Behavioral Models and V2X Communication Simulation and Case Study. 2019 IEEE Intelligent Transportation Systems Conference (ITSC),
Zhang, L., Gou, J., Liu, K., McHale, G., Ghaman, R., & Li, L. (2009, 20-23 Sept. 2009). Simulation Modeling and Application with Emergency Vehicle Presence in CORSIM. 2009 IEEE 70th Vehicular Technology Conference Fall,
Zhao, J., Guo, Y., & Duan, X. (2017). Dynamic path planning of emergency vehicles based on travel time prediction. Journal of advanced transportation, 2017.
134

APPENDIX A. EVALUATION OF EVP ACTIVATION STRATEGIES
Overview Among many potential applications of CV technology, improved Emergency Vehicle Preemption (EVP) is one that is ripe for implementation (How a Connected Emergency Vehicle Preemption System Works, 2019; Noori et al., 2016; Wang et al., 2013) as it targets a specific set of vehicles and does not require a substantial penetration of CV technology within the general traffic stream. Without CV technology, the reported benefits of EVP have been somewhat limited, especially under congested roadway conditions (Hong et al., 2012; Kwon et al., 2003; Wang et al., 2013). Many existing systems are constrained by a line-of-sight requirement between a transmitter beacon on the Emergency Response Vehicle (ERV) and a preemption request receiver at the traffic signal. While some recent deployments have integrated GPS-based triggering mechanisms for preemption (Eliminator, 2021; Emtrac; Information; Technologies, 2021) to help address this constraint, CV technology offers a potentially seamless method of integrating live ERV position data, real-time traffic data and signal status from multiple intersections to improve EVP performance at the corridor level and potentially offer ERVs route specific free-flow paths through multiple signalized intersections.
While anticipatory route clearance-based methodologies have been proposed before (Eltayeb et al., 2013; Wang et al., 2013), and have seen limited implementation using GPS and cellular-phone based technologies, there is currently very limited research on the potential effectiveness of such systems, one of the primary focuses of this project.
135

Evaluate the potential impact of CV-based EVP systems, a new Dynamic Preemption (DP) EVP strategy was developed that would take advantage of the new data streams provided by a significant penetration of CV technologies within the traffic stream. Since actual CV data streams are not currently available, the presence of such high quality of real-time traffic information (by CV enabled infrastructure) was simulated using a calibrated micro-simulation environment capable of providing simulation result in "realsimulation-time". In this document, the steps involved development of the evaluation testbed and the procedures used to evaluate the efficiency of the strategy under simulated "real world" conditions including a comparison of the proposed algorithm's performance to a traditional check-in/check-out-based EVP strategy. Both the reduction in ERV traveltime as well as potential impacts on the non-ERV vehicles were considered in this evaluation.
The general assumption with most EVP evaluations is that the EVP strategy that minimizes emergency vehicle travel time should be selected. However, consideration should be given to the possibility that strategies with similar ERV performance may have different impacts on the general traffic. To this end, two different strategies for returning the signal operations to normal operation after an EVP actuation were also evaluated by comparing the effect of the EVP event on the travel times of mainline and cross-streets vehicles.
136

Prior Research The MUTCD (2009) defines traffic signal preemption as the change in operation of a traffic signal from normal mode to a special control mode. In the case of EVP, the primary objective is to provide green indications along the path of firetrucks, ambulances, etc.. EVP provides right-of-way to the ERV, minimizing the delay in reaching the incident location and ensuring a safe and clear pathway for the ERV (WSDOT 2019), (Shaaban, Khan et al. 2019). The EVP triggering message can be conveyed to the signal cabinet through a range of methods including vehicle/driver indications (e.g., by use of strobe, siren, pushing buttons, etc.) or the infrastructure could sense ERVs through special pavement loops, radio transmission, or other vehicle to infrastructure (V2I) technologies (USDOT).
The preemption process involves two transition phases, one leading to the preemption state, and the other acting to restore normal signal operations. The Traffic Signal Timing Manual (STM) (USDOT) requires that neither the yellow signal nor all-red time intervals be omitted or shortened. Additional requirements relate to pedestrian timing constraints, returning to a red indication, allowable indication transitions, and accounting for multiple preemption requests.
Several studies evaluating EVP also considered the delays experienced by non-ERVs on opposing approaches in addition to potential ERV travel time reductions. A study by Nelson and Bullock (Nelson & Bullock, 2000) focused on investigating the effect of preemption calls on closely spaced intersections. This case study was performed in a simulated environment by linking a model built in TSIS/CORSIM to the signal
137

controllers using Hardware in the Loop Simulation (HILS) technology. Four intersections along SR-26 in Lafayette, Indiana were simulated, incorporating seven potential preemption paths with one-to-three preemption calls along each path. At each intersection, preemption calls were made at a predetermined fixed distance of the ERV from the stop-bar. Three algorithms were considered for the phase transition: a smooth transition (lengthen or shorten local cycle by up to a select maximum percentage), add time transition (local cycle may be lengthened only), and a dwell transition (controller rests in coordinated phase). It was found that the smooth transition algorithm worked best in most cases, with the level of slack time in the cycle also being an important factor. While studying the effect of ERVs, the study found that for both arterials and side streets, having a single preemption call in the simulation period had little to no effect on the overall travel time and delay.
A study evaluating the disruption of coordinated signals, using microscopic simulation models based on multiple locations in New York City, observed that the EVP related disruption took a maximum of four signal cycles to recover (Ten et al., 2003). Another study used a MATLAB simulation to study the effectiveness of two different EVP control strategies. The study suggested the use of a predefined "notification time period", designing an algorithm to minimize the ERV travel time while also minimizing the adverse effect of preemption on the side streets (Qin & Khan, 2012). There were several other studies (McHale & Collura, 2003; Yun et al., 2012) that explored this tradeoff as well, highlighting that ideally a balance needs to be maintained to manage delays for movements that conflict with the ERV path. Among the more recent EVP research efforts, (Homaei et al., 2015) demonstrated the use of fuzzy logic to select the preemption phase and extend the green time, based on demand and queue length in a V2V and V2I environment. Other studies approached this problem from the network path perspective,
138

as a route planning problem (Gedawy et al., 2008; Saravanan et al., 2017; Zhao et al., 2017), where the vehicle responds to the state of the system rather than the system responding to the needs of the vehicle. While the opportunity for optimization was relatively limited in a network where the demands are reasonably balanced, there are opportunities for minimizing the impacts to each approach by adopting a combination of EVP and route optimization.
Task Description This task involved developing and evaluating an EVP strategy to dynamically adjust the preemption trigger time to simultaneously minimize ERV delay and disruption to other traffic. In the absence of real-world CV data in the study corridor, the study created simulated data streams from a vehicle microsimulation (using VISSIM) to replicate data streams expected from future CV data. This approach was chosen to allow the rapid evaluation of multiple potential scenarios facilitated by the simulation environment.
As described above, the existing literature shows the necessity to evaluate the impact of EVP on non-ERV as well as ERV travel times. Therefore the evaluation of the dynamic preemption EVP strategies under this task focused on the ability of given strategy in providing the shortest travel-time for an ERV while limiting the delays experienced by the general traffic. The environment chosen for evaluation the dynamic preemption EVP strategies in this task consisted of a calibrated microsimulation of a medium-sized (25 intersections) real-world corridor using preemption on 8 contiguous intersections specifically focusing on EVP impacts in a coordinated system. These evaluations focused
139

on impacts during the PM peak hour, typically the most congested period for this corridor.
This study did not consider the technical aspects of CV implementation such as the range and accuracy of CV equipment. In the evaluation, the following assumptions were made relating to CV infrastructure:
CVs are present in sufficient numbers to yield a reasonably accurate estimate of the queue lengths. Existing studies have established that such estimation is feasible (Comert & Cetin, 2021; Li et al., 2013; Tiaprasert et al., 2015; Wang et al., 2021).
The path of the ERV is known and the ERV can transmit its position in real time.
The signal controllers are connected to a control center that can make signal phase change requests and signal states are known in real time.
Study Location The simulation test bed models a 6.2 mile stretch along the Peachtree Industrial Boulevard (PIB) corridor from Holcomb Bridge Rd at the south-west end to Pleasant Hill Rd on the north-east end, in Norcross, Georgia. This model includes 25 intersections on and around PIB. The layout of the network in PTV VISSIM (PTV, 2021b), and the network extents in satellite view, are shown in Figure 31 (a) & (b), respectively. For consistency in naming convention, PIB direction of travel labeled as North / South throughout the length of the corridor. The cross-street approaches are defined as
140

Eastbound (EB) and Westbound (WB). For this model, the system entities, consisting of the network geometry and the signal-heads, were built based on satellite imagery from OpenStreetMapsTM (OpenStreetMap).
Figure 31. Photos. Case Study Network of Peachtree Industrial Boulevard: (a) VISSIM Simulation
Model, (b) Satellite View by Google MapsTM (Google, 2021). Data Sources A pre-COVID pandemic time period was chosen for the input data to ensure that the traffic and signal plans in the model represented typical traffic operations. Signal control data was based on a non-holiday weekday: Tuesday, October 01, 2019. Signal plan information was obtained directly from the field controllers, representing the active plans. The signals along the corridor are semi-actuated coordinated, with a 160 second cycle. For simulation model calibration, a comprehensive volume study was not available for this corridor. Thus, for the major and minor road approach volumes, data was assimilated from multiple sources including short-term historical turn-volume count data, counts obtained from post-processing stop-bar presence detector activations, recent traffic
141

studies on the corridor, available Automated Traffic Signal Performance Measures (ATSPM), etc.. The signals on this corridor are connected to a central server where the high-resolution signal phase and timing (SPaT) data, as well as detection data, is archived. The archived data had vehicle on-off pulse information corresponding to inductive loop detectors upstream of the stop-bar for the major road through lanes. The pulse data was post-processed to generate vehicle detection data for additional volume calibration. Finally, volume balancing and volume constraint computations based on signal cycle allowance and roadway geometry were used to generate estimates to fill volume gaps or inconsistencies in available data.
Simulation Model A microscopic simulation model built using PTV's VISSIM version 2021 service pack 10(PTV, 2021b) was used for this study. VISSIM's Ring Barrier Controller (RBC) addon module was used to simulate the signal controllers and preemption strategies. The replicate runs are partially automated by using Python 3.7 (Python, 2018) scripts to drive VISSIM using its Component Object Model (COM) interface. For each simulation run the network is initially loaded with 50% of the target volume for the first 15 simulation minutes. Then the volume is raised to 100% for the next 75 simulation minutes. Effectively, out of the total 90-minute runtime, the first 30 minutes are used for model initialization, and only the last 60 minutes are used for collecting data to generate performance metrics corresponding to the PM peak hour (5 pm to 6 pm). In each simulation run, a single preemption event is modeled to ensure complete independence of the results related to each actuation.
142

ERV Behavior In addition to its default vehicle types VISSIM enables the incorporation of external vehicle models that reflect operational characteristics unique to a given vehicle type. One such model was used to create an ERV vehicle class, with specific features corresponding to a firetruck. The 3D model and characteristics were obtained from the VISSIM website (PTV, 2021a), and the speed and acceleration characteristics of the vehicle were selected to mimic a typical Heavy Goods Vehicle (HGV) in VISSIM. For this study, an ERV must pass through an intersection during a green, using lanes in the correct direction of travel. That is, the simulated ERV behavior does not allow for running a red or passing through the intersection in the lanes of the opposing traffic.
Model Calibration and Validation Model calibration, i.e., adjusting model parameters to maximize the agreement of the model behavior to field observations (Trucano et al., 2006), is an essential step to ensure that the model accurately represents field conditions. While "fine-tuning" selected key parameters of the model to mimic real traffic in the network is necessary, it is not always possible to match the traffic vehicle-to-vehicle. Nor is such matching desirable as this process may lead to overfitting the model, negatively affecting the robustness, translatability, and generalizability of the results.
In this study, the calibration effort ensured that the model sufficiently reflected the field conditions, considering both mid-block free flow speeds and saturation headways departing a signal. Validation tests with travel-time as the performance metric were
143

performed to confirm the sufficiency of the calibration. Model calibration and validation are described in the following sections.
Model Calibration I: Speed Free-flow speed is one the key-features defining the traffic flow in a network. GPS data collected on the corridor was used to determine the free-flow speed on the major road. GPS data collection equipment was deployed on 17 fire station vehicles that included firetrucks, ambulances, and ladder-trucks. The 17 vehicles were from six fire stations adjacent to the corridor. The data was collected over a 6-month period from June to December 2019. The dataset provided high resolution (two-second updates) probe vehicle data in and around the PIB network. The data was combined with the emergency response logs to separate trips where the vehicle was in emergency-response mode versus a return-trip. Only return-trips in "Flashers OFF" mode were used to represent driving behavior under normal conditions and to develop an initial speed distribution. This speed distribution underwent a deconvolution process to extract the data-points representing the free-flow speed in the network. The deconvolution process followed the methodology developed in a previous study (Anderson et al., 2019) and allowed for the estimation of the free flow speed distribution, rather than a single free flow speed value. After the deconvolution process the corridor data points were represented as a mixture of four Gaussian distributions. Of those distributions, the distribution with Mean: 52 mph, SD: 8 mph was chosen for the free flow speed distribution. This aligns well with the overall 85th percentile corridor speed of 51 mph.
144

Model Calibration II: Headway The other parameter crucial for model calibration pertinent to this analysis, is saturation headway (Dowling et al., 2004). The count detectors on the major road approaches were used to measure headways by recording the gap between two "Entry" pulses for a detector. To reflect the saturation headway distribution, the headway measurements were limited to observations of less than 4 seconds. Similarly, 6 ft pulse detectors were placed in the VISSIM model at the corresponding locations to the field detectors to obtain the corresponding simulated headway distribution.
To calibrate the VISSIM headways against those from the field detectors, the carfollowing parameters under VISSIM's Wiedemann 74 model were modified (Wiedemann, 1974). Two key parameters of Wiedemann 74 that dictate the headways are an additive part and a multiplicative factor of the safety distance. The additive part dictates the average headway, and the multiplicative part dictates the spread of the headway distribution. Four random seeds of VISSIM simulation results were compared to four weekdays of field detector data. For example, as shown in Figure 32, approximately 1175 data-points per VISSIM simulation were compared to approximately 1400 datapoints of field detector data at the NB approach of the PIB and Medlock Bridge Road intersection. The average headways (as depicted by the dotted vertical lines) for the field (labeled as MaxView) and VISSIM after calibration lie very close to each other, in the range of 2.3-2.4 seconds. The Cumulative Distribution Function (CDF) lines, shown as the orange and blue monotonically increasing lines, are also similar. A two-sample Kolmogorov Smirnov (KS) test, a non-parametric statistical hypothesis test, performed
145

on the two distributions concluded that the null hypothesis "the headway data sets come from the same distribution" cannot be rejected at a 90 percent confidence level. The calibrated VISSIM model has additive and multiplicative parts of the safety distance as 2.5 and 5.5 respectively (whereas the software defaults are 2 and 3 respectively).
Figure 32. Graph. Calibration Results: Headway Distribution for Northbound (Obenberger & Collura) Movement at PIB@Medlock bridge Rd: MaxView vs VISSIM
As a final calibration step, minor changes were made to the signal timing splits and vehicle extension timers at several intersections to better serve the synthesized traffic volumes.
146

Model Validation With travel time For model validation, travel-time was the primary reported performance metric, although volumes were also confirmed to match expected field conditions. The traffic on the PIB corridor is directional with the PM peak traffic direction being NB. The evaluation results discussed later will focus primarily on this NB direction of travel, although impacts on SB traffic were also evaluated.
Checkpoints were placed at two intermediate points on the NB PIB route to divide the corridor to ensure that sufficient complete vehicle traces were captured for each segment. The average travel time at the checkpoints were summed to determine a total average travel time for the entire 6.2 mile stretch of the NB-through route along the PIB. As a benchmark, weekday travel time data from (a) the Regional Integrated Transportation Information System ((RITIS), 2021) and (b) Google MapsTM (Google, 2021) [recorded at 5-min intervals during the 5-6 PM period on July 16, 2021] were used. The average RITIS travel time was 960 seconds, while Google MapsTM travel times averaged 825 seconds. The average travel time derived from the model was 857 seconds. As per criteria set by Federal Highway Authority (FHWA)'s Traffic Analysis Toolbox (Dowling et al., 2004), the simulated travel time should be within 15% of the field. The travel time in VISSIM lies within 15 percent of either of the data sources and satisfies the suggested validation criteria.
147

Experiment Approach Once the baseline model calibration was completed, experiments were undertaken to study the impact of preemption on ERV and non-ERV (i.e., other vehicles in the network) performance. The experiments studied the impact of EVP during the PM peak, for an ERV with a NB route along the corridor. The EVP impact was evaluated with respect to the entry transition of the signal from normal to preemption operation and the exit transition from preemption back to normal operation. Two algorithms for entry transition into preemption are considered: 1) traditional check-in check-out (CI-CO) with a preempt trigger (i.e., ERV detector in the roadway) set a fixed distance from the intersection, and 2) a dynamic preemption call placement based on the required clearance time of the estimated number of vehicles between the ERV and the signal. Additionally, two exit transition out of preemption to normal operations are also considered: 1) a normal exit with service of the phase following the preemption phase as identified in the dual ring control, and 2) an in-step exit where the preempt exits into the coordination pattern of the signal cycle. Performance metric data were collected for the ERV traveltime, travel-time of non-ERV passenger cars on the mainline, and travel-time of nonERV on the side-street for each simulation run. Additional details on the ERV route, and entry and exit transitions are provided in the next sections.
ERV Route To ensure that the effect of the signal coordination, or rather the disruption thereof, was examined in detail, the ERV route was chosen to pass through multiple intersections along the coordination path. The field ERV GPS data collected as part of this effort was
148

used to observe historical travel patterns and ensure that the path chosen was representative. The selected path is shown in Figure 33(a), (b), and (c) as a series of historical GPS points, on a map, and as part of the model network in VISSIM, respectively.

(a)

(b)

149

(c)
Figure 33. Photos. The Mainline Route chosen for the Study (a) Field ERV GPS data along the
mainline overlaid on OpenStreetMapsTM (OpenStreetMap), (b) Mainline ERV Route in Google
MapsTM, (c) Mainline ERV Route Static Routing Decision in VISSIM
Additionally, the time at which an ERV makes a preemption request, relative to the local cycle and coordination plan, was considered to likely affect the impact of preemption on the traffic. In first set of experimental results presented this effect is explored by introducing an ERV into the network at different simulation times to place preemption calls at different points in the signal cycle of the first intersection (PIB at Medlock Bridge Road) of the ERV route.
Entry Transition Two entry transitions are tested, a traditional check-in check-out (CI-CO) preemption strategy (i.e., the preemption call is placed a fixed distance from the intersection) and a customized dynamic preemption (DP) strategy. In the traditional CI-CO strategy, a
150

check-in detector is placed a fixed distance from the intersection. When the ERV reaches the detector, a preemption call is placed. The call remains active until the vehicle crosses a check-out detector (or times out), which is generally placed immediately downstream of the intersection. One drawback of this setup is experienced when an ERV enters the back of a queue that extends past the check-in detector. In this situation, the preemption call is not placed until the queue is sufficiently processed that the ERV reaches the detector. Additionally, this method is insensitive to the number of vehicles queued between the ERV and intersection, likely resulting in inefficient performance.
Dynamic Preemption (DP) Algorithm To improve the performance of CI-CO preemption, this study leveraged the potential
of real-time field detection and ERV CV data, developing a DP algorithm. In this algorithm it is assumed the ERV is continuously updating its position data. However, other traffic is assumed not to do so (i.e., they are not CVs) allowing for earlier field implementation of the proposed method. In this algorithm, the queue-length on each approach of the ERV path is monitored. Based on the queue lengths, the time of the preemption trigger is set to try and ensure that the ERV goes through the intersection without (or with minimal) reduction in speed. At a given approach a sufficient time must be allocated for the vehicles in the queue, as well as those in between the end of the queue and the ERV, to clear prior to the ERV arrival. Some transition time must also be allocated for the signal controller to serve the current phase yellow, red-clearance, and any necessary in-progress pedestrian-walk phases.
151

The primary assumption for the algorithm is that the number vehicles in the approach queue is same as the number of moving vehicles between the back of the queue and the approaching ERV. While recognized as a very rough approximation, this assumption attempts to account for field data limitations, where active counts of moving vehicles on the roadway is unlikely to be available in the short term. While it is possible within the simulation environment to determine the number of vehicles between the existing queue and the approaching ERV, utilizing this data would reduce the transferability of the method to the field. Efforts are currently underway that are exploring as to how to improve this estimate using likely available field data. For example, the queue length and the number of non-queued vehicles ahead of the ERV can, in principle, be estimated based on real-time detector data from the current and upstream intersections, or from the location information in the Basic Safety Message from CVs (with sufficient penetration of CVs). However, this estimation is a non-trivial problem, and has been studied by other researchers (Li et al., 2013; Tiaprasert et al., 2015).
In the current algorithm a preemption decision may be made as early as when the ERV begins to approach the upstream intersection for the subject approach. Thus, it is assumed that the ERV route is known at least two intersections in advance. The preempt call time is calculated as follows: If "n" cars are present in the queue, assuming that the headway is 2 seconds and an additional reaction time (i.e., start-up lost time) of 4 seconds, the time taken to clear that queue would be (4+ 2*n) seconds. To clear the "n" vehicles between the end of the queue and the ERV, an additional 2*n seconds are required. For the transition from current signal state to the preempt phase, an additional 5 seconds is added.
152

This results in an overall total of (9+4n) seconds for the advance placement of the EVP call prior to the ERV reaching the intersection. Therefore, in a corridor with a free flow speed of "v" ft/sec, the preemption will be triggered by the ERV at a distance of (9+4n) *v feet from the intersection, when the intersection in question has a queue length of n cars.
Exit Transition While the entry transition has a significant impact on the ERV's travel-time, the traveltime of the other vehicles are affected by the exit transition as well as the entry transition. To explore the impact of the transitions on the non-ERVs, the experiment was performed with the two different exit transitions to help investigate if there are any discernable impacts on (a) the ERV, (b) the mainline traffic, and (c) the cross-street traffic, with the different strategies.
In the first scenario, the controller serves a defined set of exit phases or the next phase immediately after the preempt-phases, if no exit phase is specified. If a coordination plan is in place the controller (the RBC in VISSIM in this study) would then need to adjust the local cycle and splits to transition back into coordination. This is termed as normal exit in the VISSIM RBC. (Normal exit was used for the entry study discussed above.)
The second exit transition considered is termed in-step, in which the preempt exits into the coordination pattern of the signal cycle (PTV, 2014). In this exit transition, the signal control will exit into the current phase that would be running according to the local cycle, if sufficient time exists to serve the minimum green time. If sufficient time does not exist,
153

it would exit into the next phase in the cycle if a call exists. Finally, it will exit into the coordinated phase if the prior conditions are not satisfied. Underlying this logic is the assumption that no additional transition will be required to return to coordination. In general, the in-step exit favors the main-line coordinated phase, while the normal exits seek to lessen the green time potentially taken from the non-coordinated movements.
There are previous studies that have investigated the impact of exit strategies for preemption (Mittal, 2002; Obenberger & Collura, 2001). This effort builds on these foundations by exploring the exits strategy impact on non-ERV traffic when combined with the traditional CI-CO entry transitions as well as the developed DP method.
Summary To summarize, the analysis consists of three experiments, each building on the
prior: 1. First, the impact of the time of arrival of the ERV during the cycle was explored. This investigation uses the DP algorithm with normal exit. 2. Second, the performance of Dynamic Preemption vs Check-In-Check-Out was compared. The comparison uses normal exit in both cases and, based on the findings of experiment 1, includes variations in ERV network entry time with respect to the signal cycle. 3. Third, a comparison of normal and in-step exit transitions was conducted for the DP model, again considering variations in ERV network arrival time.
154

For all experiments ERV and non-ERV travel times are considered along the mainline as well as non-ERV on the cross streets.
Results Experiment 1 - Impact of ERV arrival-time with respect to signal cycle on traveltime It is expected that the time at which an ERV makes a preemption request, relative to the local cycle and coordination plan, is likely to affect the entry into and exit from preemption and thereby affect the impact of preemption on the background traffic. As stated, this effect is explored by introducing an ERV into the network at different simulation times, resulting in ERV arrival times throughout the signal cycle of the first intersection (PIB at Medlock Bridge Road) in the route given in Figure 33 (a, b, c).
The PM cycle length for all intersections within the corridor is 160 seconds. Thus, to reflect a cross section of possible arrival times 32 different scenarios are created, with successive five seconds increments for the time of introduction of the ERV into the network. To account for stochastic variability, ten replicate runs per scenario are performed using ten random seeds. Thus, a total of 320 simulation runs (32x10) were conducted in the study of ERV arrival time. For these runs, the DP algorithm was used with normal exit transition.
In the travel time comparisons shown below, a hybrid boxplot was used to display the results (see Figure 34-Figure 39). In this approach, for the replicate trials of each scenario (ten per scenario in this experiment), the red square dots represent the mean, and the top
155

and bottom of the solid box represent the 75th percentile and 25th percentile, respectively. The difference between the 25th and 75th percentiles is the inter-quartile range (IQR). IQR dictates the amount of spread for the middle 50 percent of data-points around the median (Taylor, 2019). The black line drawn on the solid box is the median; the "whiskers" around the box span within 1.5*IQR of the box boundaries and the points that lie beyond the whiskers are considered as outliers (Galarnyk, 2018). Figure 34 (a) shows the effect of the staggered entries on the travel-time (y-axis) of the ERV (traveling the length of the eight-intersection route) with each box and whisker plot showing the travel time variability across the ten replicate runs for an ERV entering at the given simulation time (x-axis). In each successive entry time scenario, the ERV entrance into the network is increased by five seconds to ensure sufficient variability such that arrivals are distributed throughout the cycle. The time of the entrance into the network, and thus time of the preempt call, relative to the local cycle is shown to have an impact on the ERV travel time, with the mean travel time across scenarios ranging from 438s to 470s, while IQR ranges from 27s to 89s.
156

(a)

(b)

(c)

Figure 34.Graphs. (a) Impact of staggered ERV entry time on ERV Travel Time, (b) Finding the best cluster number for travel time distribution, (c) Visualization of 3 different clusters separating (ERV
entry time, ERV travel-time) data-points

A review on the plot leads to an intuitive implication that travel times from ERV entry of 3522.5s to 3587.5s yield more erratic median travel times and lesser IQR compared to the entry time 3592.5s to 3677.5s. Hence it is important that further investigation is made into the distribution of the ERV travel-time as a function of the entry of vehicles into the

157

network. An approach to doing so was explored by Anderson et al. (Anderson et al., 2019), who demonstrated a way of segregating data that follow different distributions from a composite dataset to traffic analysis on the individual separate data subsets, or "clusters", by formal definition. The same approach was previously used to get the freeflow speed distribution used in this study. A Gaussian Mixture model (GMM) was fitted to the travel-time data and a cluster analysis was performed. It is important that the optimum number of clusters are used for this exercise. Without this step, there is possibility of over-splitting or under-splitting the dataset giving us results that does not provide the most amount of information (Hayasaka). One of the best ways to find the optimal number of clusters for a GMM is finding the Silhouette scores for different cluster numbers. Silhouette scores determine how compact and well separated the clusters are and the higher the score is the better it is to use the number of clusters indicated by the score (Rousseeuw, 1987). Figure 34 (b) implies using three clusters would be optimal. After breaking the data-set into 3 clusters as depicted by Figure 34(c), it is observed that the ERV travel-times resulting with varying ERV entry-time into the network can be roughly broken into two different clusters, the green cluster between entry times 3522.5 s to 3592.5s ERV entry times and the [clusters in violet and clusters in yellow] together from ERV entry-time 3597.5 s to 3677.5 s. This closely aligns with the initial intuition gathered from Figure 34(a). This strongly suggests that variation in the cycle entry-points has a significant impact on the ERV travel-time distribution.
158

Figure 35(a) & (b) shows a similar effect on the side-street through movement travel times at the first preemption-activated intersection of PIB at the Medlock Bridge Road intersection. The impact is measured from the preemption event to two cycles after the end of preempt. Mean side-street travel time for through moving vehicles ranged from 136s to 193s and from 124s to 174s on the EB and WB through movements, respectively. In addition, the IQR for travel time ranged from 72s to 127s and from 80s to 144s on the EB and WB through movements, respectively. Given the range of variability seen for all movements in traffic, all subsequent experiments include using a range of times for the infusion of the ERV into the network, uniformly covering the cycle length of the signals of the corridor (i.e., 160s).

(a)

(b)

Figure 35. Graphs. Variation in Travel Time for side-street through movement (PIB @ Medlock Bridge Road) for (a) EB Through and (b) WB Through with Different ERV Arrival Times.

Experiment 2: Dynamic Preemption vs Check-In Check Out

As shown in Figure 33(b), for these simulations the ERV entered the Southern part of the network, upstream of PIB @ Tech Parkway South and traveled NB through eight
159

intersections, from Medlock Bridge Road to Howell Ferry Road where it made a right turn and exited the corridor. The trajectory plot for this route was previously shown in Figure A1 These simulations were designed to examine two potential issues associated with CI-CO EVP activation: 1) if the back of the signal queue extends beyond the checkin detector the ERV does not trigger the preemption call until it advances in the queue to the detector location, and 2) in CI-CO activation the call is set at a fixed distance without consideration of the real-time traffic conditions. Under these circumstances, the DP approach might be expected to improve performance by making the preemption request further upstream thus increasing the likelihood of a successful queue flush prior to the ERV arrival at the intersection box. As with the prior experiment, there are 32 runs of ERV network entry samples for each random seed, thus 320 total runs out of the 10 random seeds.
Figure 36 provides a visualization of the difference between these methods. The color in the ERV trajectory represents the GREEN/AMBER/RED signal state of the next downstream intersection at the corresponding time. The blue line running in parallel to the trajectory represents the duration of an active preemption call by the ERV while traveling along its trajectory. Figure 36(a) and (b) present an example scenario, comparing an ERV entry into the network at the same simulated time, for CI-CO and DP. Figure 36(a) has preemption enabled with normal exit and a CI-CO implementation, where the detector is placed on an approach 1000 ft upstream of the intersection or, where 1000ft is not available, immediately after the upstream intersection. (b), with the same entry time as Figure 36(a), depicts the ERV trajectory using DP to place preemption
160

calls and normal exit. Within this example the queue length drawback of CI-CO is evident, particularly at closely spaced intersections. In the CI-CO strategy the ERV is unable to place the call sufficiently early to allow for the queue to clear, but rather being delayed in the queue while the downstream vehicles clear. The DP approach was shown clear the queue more consistently prior to the ERV arrival, resulting in a significant reduction in delay.

(a)

(b)

Figure 36. Graphs. ERV Trajectory in the PIB VISSIM network: preemption enabled: (a) CI-CO, (b) DP

The impact of this difference in preempt methods may be seen in the impact to ERV and non-ERV travel times shown in Figure 37. For the ERV travel time, Figure 37(a), averaged across all arrival times, the CI-CO provides an approximately 73s advantage over no-preempt while DP provides a 142s advantage. The difference in the ERV travel time between Preempt disabled and CI-CO , as well as between CI-CO and DP, were found to be statistically significant, based on a non-parametric one-way Kruskal-Wallis H-test (McKight & Najab, 2010) at 5% significance level. This comparison provides a
161

strong demonstration of the advantages of an approach, such as DP, leveraging CV data. When considering the impact on the side-streets as reflected by Figure 37(b) within the first two signal cycle lengths after the preemption activity, it is seen that the effects on side street traffic were comparable, with the average travel time being 107s, 130s and 131s for "no preempt", CI-CO, and DP respectively. Similar trends are observed for other intersections as well. Hence the overall gain in ERV response time with negligible excess side-street traffic delay provides a strong argument for EVP in general, as well as using CV-based activation methods.
(a)
162

(b)
Figure 37. Graphs. Overall travel time under three entry-transition experimental setups: (1) Preempt Disabled, (2) Check-in Check-out (with Normal Exit), (3) Dynamic Preemption (with Normal Exit) for (a) (top) ERV through the designated route; (b) Non-ERVs at PIB@ Highwo Highwoods Center WB-Through for 2 signal cycles after EVP activity.
Experiment 3: Exit Transition Analysis The best EVP strategy would be one that would minimize ERV delay, while also minimizing the negative effects on the other traffic, especially on the cross-streets (Shaaban et al., 2019a). While the negative impacts of EVP are likely to be more pronounced on the side streets, it is possible that negative impacts may also be experienced by vehicles on the preemption approach. Therefore, this experiment considered both the preemption approach and the cross-streets. To avoid the confounding effects of turn-related delays due to mixing of through and turn movements, the measurements in this study are made for vehicles involved only in through movements for both the main-line and cross-streets.
163

As with the prior comparisons, the impact of the preemption exit treatment on travel time were evaluated by aggregating over 32 ERV network entry times distributed over the cycle at the initial intersection, with ten replications of each. For this study only DP is considered, thus three preemption strategies are tested: no-preemption, DP with normal exit transition, and DP with in-step exit transition. Thus, a total of 960 (32x10x3) simulation runs were undertaken for this experiment. For the subsequent analysis the impact of each strategy is considered over six time-intervals after the ERV first enters the network, where each interval is one cycle length, e.g., 160 seconds, allowing for observation before the preemption call is placed as well as dissipation effects.
In the analysis the interval data includes those vehicles that end their travel through the corresponding travel-time segment during that interval. That is, interval aggregations are based on the time an ERV or non-ERV leaves a segment rather than enters. Figure 38(a) summarizes the travel time variation for the ERV in the network under the different preemption scenarios. The average travel time for no-preemption, preemption enabled with normal exit, and preemption-enabled with in-step exit, are 599s, 457s and 454s, respectively, with an approximately 140s average travel time saved by both preemption exit modes. The improvement for both exit-strategies is statistically significant relative to no preemption, while there is no statistically significant difference between the two exit strategies. Another positive aspect of preemption was a reduction in the variation of travel time as reflected by the reduced IQR, seen in both exit transition strategies.
164

(a)

(b)

Figure 38. Graphs. Mainline approach preemption scenario travel times over the entire mainline
(Obenberger & Collura) ERV route: (a) ERVs, (b) Non-ERVs upstream of ERV.

Figure 38(b) shows the travel time for non-ERVs (i.e., general traffic) traveling the same route as the ERVs. Here, it is seen that there is minimal impact within the first 320 seconds of the ERV entrance into the network on the non-ERVs that are traveling the same route as the ERV, as these vehicles are sufficiently downstream of the ERV that they pass through each intersection prior to any preempt call being placed along their route.

However, a positive impact is experienced by the non-ERVs that arrive after a preempt call (or are in queue when the call is placed), either benefiting from the clearing of the
165

non-ERV downstream of the ERV or following the ERV through the intersection. Both the normal and in-step preemption strategies show significant improvement in travel time for the non-ERVs along the ERV route, after the preemption call is placed. The impact was most significant for vehicles that are in proximity of the ERV, as the DP-activated EVP design across multiple contiguous intersections creates a band of Green-signal states along the ERV trajectory. Vehicles leading the ERV benefit from the queue clearance forced into the system by EVP. Traffic immediately behind the ERV also benefits given the overall queue clearance along the main-line corridor as the ERV progresses down the corridor.
The travel-time trends were compared statistically using Kruskal-Wallis test for the medians of the travel-time populations of the different strategies. As expected, it was observed at a 5% significance level that the null-hypothesis: "the median travel-time with and without preemption are not different" could not be rejected for both Normal and InStep exit scenarios until 320s after the entry of ERV. The null hypothesis was rejected in the subsequent time-intervals.
Figure 39(a) & (b) illustrate travel times along two representative cross streets, again aggregated over the 32 arrival times, with ten replicate trials, for the three preemption strategies comparing travel times on the side streets for the two preemption strategies (CI-CO and DP) relative to a baseline of no preemption. From these plots, we see that vehicles in the first interval following the exit of the ERV from the intersection experience the most significant impacts of an EVP call. Figure 39(a) & (b) represent the
166

best and the worst case of cross street EVP-effect observed in the network respectively. As expected, at both intersections the travel time increases for the vehicles that are delayed due to a preemption call with these impacts dissipating within three signal-cycle lengths (i.e., 480 seconds) of the preemption call as demonstrated by a Kruskal-Wallis test. Within the disruption period, a slightly higher IQR was observed in the worst-case intersection for in-step preemption, possibly because the strategy involves higher GREEN time for the ERVs on average. The trend of the greatest impact occurring immediately after the preempt event and dissipating rapidly within three cycles, was consistent for all side street traffic across the corridor.

(a)

(b)

Figure 39. Graphs. Travel times for side-street trajectories post mainline approach preemption: WBThrough movement at (a) (left) PIB @South Old Peachtree Road, (b) PIB @North Berkeley Lake Road

167

Experiment 4: VISSIM-MaxTime Connection: SILS Study In this experiment, the preemption impact results evaluated using VISSIM's in-built RBC signal controller to emulate signal operations discussed in the previous sections are compared to those obtained using the MaxTime controller software in a Software-inthe-Loop (SILS) simulation. Figure 40 compares these SILS results to the equivalent RBC results for non-ERVs following the ERV. The normal exit strategy is used in the SILS implementation. The trends and absolute values are similar for the two implementations. Each method shows improvement in non-ERV travel times of vehicles upstream of the ERV (i.e., vehicles following the ERV). Non-ERVs within that same departure bin (approximately 500 seconds) receive the most benefit having followed the ERV's green wave, with the benefits dissipating as additional time passes.
RBC:

(a)

(b)

168

Figure 40. Graphs. Boxplots depicting variations in time taken to travel the entire NB ERV route for Non-ERVs upstream of the ERV: (a) with RBC, (b) with MaxTime i.e., SILS.
Figure 41 shows the impact of preemption on ERV travel times in SILS as compared to the VISSIM RBC-driven model. In SILS, the average travel-time without preemption was 614 s and with preemption, there was a 132 s drop (~21 %). This is on par with the 20% improvement seen in the RBC implementation. Figure 42(a) and (b) depict the impact of preemption call on the side-street traffic travel times for RBC and SILS. The disruptions are significant mostly in the first two signal cycle lengths. The effect dissipates rapidly after that for both implementations. Statistically supported by a KW test on travel-time, the difference was insignificant beyond three signal-cycle lengths. For MaxTime, the preemption algorithm was run with all the parameters set to their default values. Better tuning of the MaxTime preemption parameters might provide and opportunity for improvement in the preemption operations.
169

(a)

(b)

Figure 41. Graphs. Boxplots depicting variations in time taken by ERVs to travel the entire NB ERV route (a) with RBC, (b) SILS.

(a)

(b)

Figure 42. Graphs. Travel time comparison for side-street non-ERVs upstream of ERV, with varying preemption-strategy for EB-through movement on North Berkeley Lake Road at PIB: (a) with RBC,
(b) SILS.

Conclusions

This task aimed to quantify the impacts of the use of CV technology on EVP by implementing a DP architecture in a microscopic simulation model with ERV travel time as the primary metric. For the cases studied, DP with either exit transition led to an approximate 23% (~140s) reduction in average travel time for the ERV. While not as significant, it was seen that traditional CI-CO also had the potential to improve ERV travel time, with approximately half the benefit seen from the DP algorithm. In the DP scenarios, the non-ERVs sharing the same path as the ERV also received a significant
170

reduction delay as a by-product of the preemption. When considering the side-street there was a preemption-derived disruption in travel leading to increased cross street travel times. However, this effect dissipated quickly. SILS provided similar results to the VISSIM RBC simulations.
Since the normal exit transitions were observed to be slightly superior in service to the side streets during the disruption period, DP with normal exit was found to be the most favorable preemption strategy for the network and conditions studied. However, generalization of these results will require additional study of other corridors and under varying traffic conditions with other exit modes beyond those discussed here.
The observed improvements or degradation in travel time are not directly transferrable to other corridors, each corridor with unique characteristics and demands requiring analysis. However, both this and other studies have demonstrated the potential benefits of EVP strategies and the potential for additional benefits with CV implementation.
The work on this task also identified areas for potential improvements in future studies. First, a mix of the main and side-streets along the ERV route versus the mainline-only route used in the study might provide additional insight into network effects. Second, scenarios with arrival of multiple ERVs at an intersection should be explored to better understand the interaction of overlapping preemption request calls. Third, there have been several studies tackling path planning problems for EVP (Gedawy et al., 2008; Saravanan et al., 2017; Zhao et al., 2017). Since it was known a prior that the ERV will only be traveling on the mainline utilizing a designated route, this is not included in the
171

current experimental architecture. Fourth, the comparative analysis of EVP could have been extended beyond just travel time to additional metrics such as average queue-length, cycle failures on the cross streets, etc. to provide more depth to the findings. Lastly, while successful, the DP algorithm developed is a relatively simple heuristic and additional refinement of the algorithm is probably merited.
172

REFERENCES for Appendix A
(RITIS), R. I. T. I. S. (2021). Regional Integrated Transportation Information System (RITIS). Retrieved July 28 from https://www.ritis.org/traffic/
Anderson, J., Suh, W., Guin, A., Hunter, M., & Rodgers, M. O. (2019). Accounting for composite travel time distributions within a traffic stream in determining Levelof-Service. Journal of Intelligent & Fuzzy Systems, 36(2), 955-965.
Comert, G., & Cetin, M. (2021). Queue length estimation from connected vehicles with range measurement sensors at traffic signals. Applied Mathematical Modelling, 99, 418-434.
Dowling, R., Skabardonis, A., & Alexiadis, V. (2004). Traffic analysis toolbox, volume III: Guidelines for applying traffic microsimulation modeling software.
Eliminator. (2021). Preemption System - Emergency Vehicle - Transit. Collison Control Communications. Retrieved July 27 from https://collisioncontrol.net/PreemptionSystems/?gclid=Cj0KCQiApt_xBRDxARI sAAMUMu_IjJEiTe4k0tnTcFLBTB6GNLLJ1F9WstFXn0ty2tgO5SPZwVc1syoa AqMyEALw_wcB
Eltayeb, A. S., Almubarak, H. O., & Attia, T. A. (2013). A GPS based traffic light preemption control system for emergency vehicles. 2013 international conference on computing, electrical and electronic engineering (icceee),
Emtrac. EMTRAC Systems Signal Priority: Emergency Vehicle Preemption (EVP) Utilizing ITS Functionality. Retrieved July 27 from https://www.emtracsystems.com/first-response/emergency-vehicle-preemptionevp/
Galarnyk, M. (2018). Understanding Boxplots. towards data science. Retrieved July 30 from https://towardsdatascience.com/understanding-boxplots-5e2df7bcbd51
Gedawy, H. K., Dias, M., & Harras, K. (2008). Dynamic path planning and traffic light coordination for emergency vehicle routing. Comp. Sci. Dept., Carnegie Mellon Univ., Pittsburgh, Pennsylvania, USA, 189.
Google. (2021). Google Maps: Peachtree Industrial Boulevard. Retrieved July 27 from https://www.google.com/maps/dir/33.9459033,-84.2363882/34.0064816,84.1701493/@33.9949555,-84.2917807,11679m/data=!3m1!1e3!4m2!4m1!3e0 173

Hayasaka, S. (2022, Feb 11, 2022). How Many Clusters? Methods for choosing the right number of clusters. Towards Data Science. Retrieved June 01 from https://towardsdatascience.com/how-many-clusters6b3f220f0ef5#:~:text=For%20some%20clustering%20algorithms%2C%20such,i nformative%20(see%20Figure%201).
Homaei, H., Hejazi, S., & Dehghan, S. A. M. (2015). A new traffic light controller using fuzzy logic for a full single junction involving emergency vehicle preemption. Journal of Uncertain Systems, 9(1), 49-61.
Hong, K.-S., Jung, J.-H., & Ahn, G.-H. (2012). Development of the Emergency Vehicle Preemption Control System Based on UTIS. The Journal of The Korea Institute of Intelligent Transport Systems, 11(2), 39-47.
How a Connected Emergency Vehicle Preemption System Works. (2019). applied information. Retrieved Nov 05 from https://appinfoinc.com/how-preemptionsystem-works/
Information, A. Emergency Vehicle Preemption and Priority System: Applied Information. Retrieved July 27 from appinfoinc.com/solutions/preemptionpriority/.
Kwon, E., Kim, S., & Betts, R. (2003). Route-based dynamic preemption of traffic signals for emergency vehicle operations. Transportation Research Board 82nd Annual MeetingTransportation Research Board,
Li, J.-Q., Zhou, K., Shladover, S. E., & Skabardonis, A. (2013). Estimating queue length under connected vehicle technology: Using probe vehicle, loop detector, and fused data. Transportation Research Record, 2356(1), 17-22.
McHale, G. M., & Collura, J. (2003). Improving emergency vehicle traffic signal priority system assessment methodologies. Transportation Research Board 82nd Annual MeetingTransportation Research Board,
McKight, P. E., & Najab, J. (2010). Kruskal-wallis test. The corsini encyclopedia of psychology, 1-1.
Mittal, M. S. (2002). Assessing the Performance of an Emergency Vehicle Preemption System: A Case Study on US 1 in Fairfax County, Virginia Virginia Tech].
174

Nelson, E. J., & Bullock, D. (2000). Impact of emergency vehicle preemption on signalized corridor operation: An evaluation. Transportation Research Record, 1727(1), 1-11.
Noori, H., Fu, L., & Shiravi, S. (2016). A connected vehicle based traffic signal control strategy for emergency vehicle preemption. Transportation Research Board 95th Annual Meeting,
Obenberger, J., & Collura, J. (2001). Transition strategies to exit preemption control: State-of-the-practice assessment. Transportation Research Record, 1748(1), 7279.
OpenStreetMap. OpenStreetMap: Peachtree Industrial Boulevard Corridor. OpenStreetMap. Retrieved January 17 from https://www.openstreetmap.org/directions?engine=fossgis_osrm_car&route=33.9 4488%2C-84.23857%3B34.01961%2C-84.16084#map=13/33.9823/-84.1997
PTV, A. (2014). PTV Vissim 2021: RBC user manual. PTV AG: Karlsruhe, Germany.
PTV, A. (2021a). PTV Vissim- 3D Vehicle Models. PTV AG: Karlsruhe, Germany. Retrieved July 28 from https://company.ptvgroup.com/en-us/ptv-vissim-3dvehicle-models
PTV, A. (2021b). PTV Vissim 2021 user manual. PTV AG: Karlsruhe, Germany.
Python. (2018). Python 3.7.0 Release. Retrieved June 25, 2020 from https://www.python.org/downloads/release/python-370/
Qin, X., & Khan, A. M. (2012). Control strategies of traffic signal timing transition for emergency vehicle preemption. Transportation research part C: emerging technologies, 25, 1-17.
Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20, 53-65.
Saravanan, R., Paul, P. V., Nivedha, M., & Nirosha, B. (2017). Optimal path planning and traffic signal control system for emergency vehicle--A survey. 2017 International Conference on Computation of Power, Energy Information and Commuincation (ICCPEIC),
175

Shaaban, K., Khan, M. A., Hamila, R., & Ghanim, M. (2019a). A strategy for emergency vehicle preemption and route selection. Arabian Journal for Science and Engineering, 44(10), 8905-8913.
Taylor, C. (2019). Understanding the Interquartile Range in Statistics. ThoughtCo. Retrieved August 1 from https://www.thoughtco.com/what-is-the-interquartilerange-3126245
Technologies, G. T. (2021). Central Management Software (CMS) for Traffic Signal Priority Control. Global Traffic Technologies. Retrieved July 27 from www.gtt.com/project/central-management-software-cms-for-traffic-signalpriority-control/
Ten, H., Qi, Y., Falcocchio, J. C., Kim, K.-B., Patel, R., & Athanailos, E. (2003). Simulation testing of adaptive control, bus priority and emergency vehicle preemption in New York City. Transportation Research Board 82nd Annual MeetingTransportation Research Board,
Tiaprasert, K., Zhang, Y., Wang, X. B., & Zeng, X. (2015). Queue length estimation using connected vehicle technology for adaptive signal control. IEEE Transactions on Intelligent Transportation Systems, 16(4), 2129-2140.
Trucano, T. G., Swiler, L. P., Igusa, T., Oberkampf, W. L., & Pilch, M. (2006). Calibration, validation, and sensitivity analysis: What's what. Reliability Engineering & System Safety, 91(10-11), 1331-1357.
Wang, Y., Wu, Z., Yang, X., & Huang, L. (2013). Design and implementation of an emergency vehicle signal preemption system based on cooperative vehicleinfrastructure technology. Advances in Mechanical Engineering, 5, 834976.
Wang, Y., Yao, Z., Cheng, Y., Jiang, Y., & Ran, B. (2021). Kalman filtering method for real-time queue length estimation in a connected vehicle environment. Transportation Research Record, 2675(10), 578-589.
Wiedemann, R. (1974). Simulation des Strassenverkehrsflusses.
Yun, I., Park, B. B., Lee, C. K., & Oh, Y. T. (2012). Comparison of emergency vehicle preemption methods using a hardware-in-the-loop simulation. KSCE Journal of Civil Engineering, 16(6), 1057-1063.
Zhao, J., Guo, Y., & Duan, X. (2017). Dynamic path planning of emergency vehicles based on travel time prediction. Journal of advanced transportation, 2017. 176

APPENDIX B. EVALUATION OF MACHINE LEARNING BASED EVP ACTUATION STRATEGIES
Drawbacks of "Dynamic Preemption" Architecture and Possible Solutions The DP algorithm discussed in Appendix A assumes the real-time availability of queue lengths and makes a crude assumption for the time required to clear the vehicles between the ERV and the back of the intersection queue. Figure 43 shows the average ERV speed at each intersection along the PIB test corridor, determined at the exit of the upstream intersection to the stop-bar of the subject intersection for the initial random seed of the simulations. The dashed line represents the average free flow speed for an ERV (i.e., 53 mph). As seen in the figure, the proposed DP algorithm, even with the assumption of ready availability of real-time queue length, does not achieve an ERV speed close to free flow speed. In fact, in most roadway sections, the DP algorithm fails to reach a speed of 40 mph on a 45-mph corridor and there is clearly room for improvement.
177

Figure 43. Graph. Average Speed [mph] for ERV at each intersection: Random Seed 1
Importantly, accurate real-time queue length data is not widely available given the current and near-term state of deployment of CV technology. To increase the robustness and the usability of the algorithm, this task aimed to develop procedures to improve preemption timing by using the available real-time attributes of the corridor and individual signals.
For example, consider the scenario in Figure 44. The example consists of two intersections, Intersection A and Intersection B. The ERV is upstream of Intersection B and needs to traverse both Intersections B and A. At Intersection A the ERV approach is currently RED, and at Intersection B it is GREEN. There is a queue of length Q at the downstream intersection A and there are moving vehicles between the back of the queue and the front of the ERV. A preemption call needs to be made with sufficient time to clear the queue and the intermediary vehicles so that the ERV may reach and traverse
178

Intersection A at near free-flow speed. Based on initial intuition, the decision about preemption will depend on:
1. Queue Length Q for each lane of the subject intersection approach (A) 2. Number of Cars (marked in X) between the back of the queue and the ERV 3. Distance of the ERV from the intersection stop-bar (d) 4. States of the signals at intersections A and B Given this complete information it becomes possible to estimate the necessary green time at Intersection A to clear all queued and approach vehicles such that the ERV may traverse the intersection without slowing. This is the underlying approach of the DP method; however, requiring an assumption for the number of vehicles traveling on the roadway since that is not available as an observation from detectors in the field. Thus, under current conditions this time estimate will require a decision based on incomplete information. Such decisions are common in both transportation and other fields. In recent years substantial progress has been made in the use of Machine Learning (Mohammadi et al.) methods for assisting such decisions. This task was aimed at evaluating the potential for ML for improving timing decisions associated with EVP applications.
179

Figure 44. Illustration. Schematic Depiction of Factors Determining the need for EVP Call at a given intersection
Use of Machine Learning to Solve Transportation Problems In addition to the novel machine learning (Mohammadi et al.) and AI concepts
being created to support self-driving vehicle implementation (Chen & Huang, 2017; Grzywaczewski, 2017; Hse et al., 2019; Hossain et al., 2018; Rao & Frtunikj, 2018) Grzywaczewski A. (2017), Rao Q, Frtunikj J. et al., (2018), Hse F, Roch LM, and, Aspuru-Guzik A (2019), Hossain S, Fayjie AR, Doukhi O, Lee D-J, et al., (2018), Chen Z, Huang X, et al., (2017)) the use of ML to understand aspects of travel behavior and decisions in transportation has a long history (Arciszewski T, Khasnabis S, Hoda SK, and Ziarko W., (1994)).
This discussion focuses primarily on studies related to EVP. For example, a detection study by Islam and Abdel-Aty (2022) proposes a Support Vector Machine (SVM)
180

strategy to detect an ERV with a very high accuracy in real-time with the aid of audiodata. A recent study performed by Pant et al., (2022) tackles the arrival of ERVs in congestion scenarios. Pant et al. proposed a dynamic method that uses Kullback-Leibler (KL) divergence to measure traffic density and prepare a corridor for the EVP call. The study also uses experiments on different scenarios using the Simulation of Urban MObility (Cong Minh Ho Chi et al.) traffic simulator. Savithramma et al. (2022) tackle a problem very similar to the current effort using Convoluted Neural Network (CNN) to predict the required green time at a given signalized interchange. The study made a comparison of the trained preemption model and the controller with no preemption and found significant improvement in ERV response time with the proposed solution. A study performed by Shaaban et al. (2019) seeks the optimal point of the start of the GREEN time accounting for conflicting traffic delay. It considers an optimal path suggestion for the ERV as well as the optimal time for the EVP indication. The experimental setup in this case is a microscopic simulation model consisting of four contiguous signalized interchanges built in VISSIM based on a corridor in Doha, Qatar. The study assumed significant penetration of V2V and V2I capabilities. The effort suggests a deterministic model to pinpoint the optimal EVP indication time, while determining the optimal ERV path using Dijkstra's algorithm (Dijkstra EW (1959)). The study suggested possible extensions to the effort, such as (a) using real-time feedback from the vehicle fleet and (b) leveraging of ML techniques to handle more complicated real-world incidents. These are essential points that help form the basis of the current effort. There are several other EVP studies that leverage novel ML techniques that are tangential to the current effort. For example, (a) designing EVP in the context of uncertain road incidents (Min et al.,
181

2019) and (b) surveys tackling EVP problems based on routing and scheduling (Kamble & Kounte, 2022) that provide motivation for the algorithm development discussed here to varying degrees.
Relationship to Task 1 As discussed in Appendix A, it is imperative that a study of the impact of EVP account for the travel time on both the route of the ERV as well as the routes that are interrupted. While there have been previous studies on the evaluation of innovative algorithms to optimize EVP performance and traffic disruption, these studies are typically limited to preemption at a single intersection or a few non-contiguous intersections in a small network. The research undertaken in Task 1 of this project tackled the EVP problem through simulation of a medium sized network (25 intersections) and studied the implementation of preemption on eight contiguous intersections focusing on the challenge of mainline signal coordination disruption and side-street traffic impacts. Task 1 demonstrated how integration of real-time traffic flow information from CV technologies (OBU on the ERV and RSUs at the intersections) could feed ad hoc heuristic "Dynamic Preemption" (DP) logic to improve the movement of the ERV through the intersections. The DP study was also extended to incorporate a SILS architecture in the scenario analysis.
An existent weakness of the DP algorithm is that it assumes significant knowledge regarding queue lengths at the intersections along the ERV path, information that is likely to be difficult to obtain in the short term. In the absence of this information, this task was aimed at examining the extent to which the simulations developed in Task 1 could be
182

used to train a sophisticated ML-model to achieve similar performance to the queueinformation based DP method using only information likely to be available in the nearterm.
Research Approach Initial Proposed Methodology A supervised ML approach was chosen for the development of an optimal preemption strategy. The training data set for the ML was limited to data that may be obtained on a reasonably equipped corridor today. That is, vehicle detection information is available in real-time, the ERV has GPS equipment and has connectivity with the traffic signals directly or through a central server, and the traffic signals have remote connectivity and stream signal state (SPaT) data available.
The calibrated corridor model presented for the PIB corridor was used to train the ML model with the simulation results assumed to represent perfect knowledge of corridor conditions. However, the ML training data were limited to the data that would be available in a field environment. The development of the ML method consisted of the following steps:
4) Optimal EVP calls were developed for a number of scenarios utilizing "perfect" information, that is, full knowledge of the number and placement of vehicles based on simulated results. These preemption calls were used for the training data set. The optimal preempt call times were developed using an iterative trial and error method. 183

5) The motivation for proposing an approach that uses partial, but realistically available data in real-time is that it would mimic the current field situation of limited availability of traffic information. However, it is assumed that even with limited available real-time data, the network infrastructure can provide second-bysecond ERV trajectories using CV technology, as well as SPaT messages providing the signal and detector state. These data were used to train a supervised learning model using the solution suggested in step (1) to build a preemption decision algorithm utilizing the ERV trajectory and SPaT messages. . Different supervised learning algorithms are tested at each intersection and the best performing models were chosen.
6) The EVP recommender models formed for contiguous intersections were combined and run concurrently to create a system-level EVP recommender model. The composite system-level EVP recommender system was used to predict the best time to activate EVP at any given intersection within the ERV route, in real-time as the ERV traverses across the corridor.
All steps were initially developed in the RBC only environment with final testing in the SILS environment.
Justification for Multiple Intersection Optimization A space-time diagram of vehicles traversing the mainline of the corridor, including an ERV with preempt calls using the DP algorithm (Appendix A) is shown in Figure 45. The ERV trajectory is marked in blue, in contrast with other vehicles marked in gray. The
184

horizontal line colors in Red, Green, and Yellow represent the signal states of the signals for the through movement (i.e., preempt movement) at each intersection. The orange highlight on top of the signal lines represents the duration for which each intersection had the EVP call.
Figure 45. Graph. Trajectory of Other Vehicles around ERV [marked in blue] for one given Random Seed (1) of Simulation where ERV enters the Network at t=3587.5 s [t=0 at 4:30 PM]
For the scenario in the figure, consider North Berkeley Lake Road. The EVP call is made almost immediately when the ERV crosses the intersection two blocks upstream (i.e., Pickneyville Park), which is the earliest a call can be made under the DP strategy. As seen in the figure, this does not provide sufficient time to clear the queues at North
185

Berkeley Road. In addition, due to queue clearance at the upstream intersections resulting from EVP calls, upstream and downstream platoons begin to merge, causing more strain on the North Berkeley Road intersection, which the current EVP strategy was unable to clear out in time to avoid causing significant delay for the ERV. Hence, an EVP call at one intersection can have clear impacts on the optimal EVP call time at subsequent intersections. Thus, instead of considering the EVP calls as a series of singleintersection or two-intersection joint decisions, a more robust approach is to consider the entire network as a single system. Once the ERV enters the network, an EVP call algorithm should determine the time for EVP calls at all the intersections, potentially refining the call times as the ERV traverses the corridor. At each individual intersection level, the EVP has two goals, (a) clearing the vehicle platoons from upstream intersections that are being pushed toward the current intersection, as well as (b) clearing the queued vehicles immediately preceding the ERV arrival at the intersection.
Numerous signal optimization programs have been developed over the years to synchronize signals for optimum traffic flow and dispersion (Chang & Messer, 1991; Cohen, 1982; Shoufeng et al., 2008). The multi-intersection ML model would enable incorporating EVP into these logics.
ML-model Development The three main steps of the ML development process presented in this section are:
(1) Finding the optimum preempt solution,
186

(2) Developing the ML approach to be used, and
(3) Training the ML using the optimum solution.
In the first step, the optimal EVP decisions for each intersection are determined, accounting for network interactions and varying ERV entry times, using a binary search method involving multiple simulation runs. The aim of Step 2 is to build an intersection level supervised learning model for preemption determination that relies on traffic features (e.g., distance of ERV from given intersection, current signal state, etc.). Traffic features are chosen such that they may be realistically obtained in real-time, given the existing state of CV penetration. The ML algorithm at each intersection is executed each time step (a one-second here) determining whether to implement a preempt call at that intersection. Several supervised learning algorithms were tested. The algorithm, along with a combination of its hyper-parameters, that best fit at an intersection (i.e., best travel-time - preemption duration trade-off) was chosen for that intersection. In Step 3 the optimal solutions from Step 1 are used to study the variation in solution patterns as a function of the features.
Finding the Optimum Solution As has been observed previously, the earlier EVP is provided, the better it will be for the ERV response time. However, there is a preempt call time placement, in advance of which having EVP turned on produced little, if any, response time benefit, although it continues to increase the disbenefit to the conflicting traffic. The objective of the optimal
187

preempt call time is to find out that point at which the plateauing of EVP related benefits is observed.
To find the optimal preempt time an iterative search is performed. First the optimal time is determined for the nearest intersection upon entry of the ERV into the network. The optimal time is the minimum preemption duration that ensures the lowest possible ERV travel-time, measured from exiting the upstream intersection to the stop-bar of the given intersection. Once the optimal preemption time is determined a longer preemption duration will not lower the ERV travel time but may increase conflicting movement delays. Once the preempt call time is determined for the nearest intersection the call is assumed to be implemented and the optimal time is then found for the next downstream intersection. The preemption search for this next intersection again starts when the ERV enters the network, accounting for the time of placement of preemption of any intersections that have already been set. Thus, it is possible for a preemption call to be placed earlier for a downstream intersection than an upstream intersection. This process is continued until the optimal time is found for all intersections along the ERV path.
To implement the search strategy above an iterative bounding approach was used. The two extremes of an EVP strategy are (1) No-EVP: having no EVP, and (2) Extreme-EVP: having EVP since the entry of the ERV into the network. A search needs to be performed between these two extremes to determine the EVP call time at which no additional enhancement in ERV travel time is seen with an earlier preempt time. Hence, the intuitive assumption here is that if no preemption is provided, the travel time is expected to be the
188

longest, then as EVP duration increases travel time monotonically decreases until the minimum is found. At this minimum point, further increase in the preempt duration does not decrease the travel time. To find this point VISSIM runs are made for the two extremes (i.e., no EVP and EVP upon entry into the network) and then the subsequent trials are used to find the preempt activation time-stamp where the improvement in travel time plateaus. That process is performed using the binary search method as shown in Figure 46.
Figure 46. Matrix. Flowchart for Optimal EVP Call using VISSIM simulation iterations.
189

Binary Search Implementation on PIB Corridor
For the main-line ERV route chosen, the iterative process discussed above was implemented for the seven contiguous intersections, starting with PIB@ South Old Peachtree Road and ending in PIB@ Howell Ferry Road. With the most upstream intersection in the route, i.e., PIB@ Medlock Bridge Road being close to the entry-point of the ERV, EVP is tuned ON under all circumstances immediately upon the ERV entry into the network. Solution searching is completed for the 160 scenarios as discussed in Appendix A (i.e., the ERV introduced into the network at five second increments in a 160 second cycle length, with five replications of each entry time).
Upon execution of the Optimal EVP algorithm a solution is obtained for each combination of a simulation random seed and entry time for an ERV into the network. An example of this is demonstrated by Figure 47(a) & (b). In Figure 47(a), the blue line represents the trajectory of the ERV, with other vehicles indicated by the grey lines. The cyan horizontal line represents the time duration for which EVP call is active at the intersection. The horizontal lines above and below the cyan line represent the left and through indications of the signal respectively. Figure 47(b) demonstrates the EVP call start timings with each subsequent iteration. The variation in ERV travel time with each iteration is seen, with iteration eight providing the minimum EVP call needed to achieve the minimum ERV travel-time, i.e., all EVPs with a greater duration (i.e., iterations 2, 4, and 5) have the same ERV travel time.
190

(a)
(b)
Figure 47. Graphs. PIB@ South Old Peachtree Road: Random Seed 5: (a) Example of ERV trajectory at various EVP call start times; (b) Variation of Stop-bar to Stop-bar travel time, with
variation in the start-time of EVP call (trial numbers labelling each data-point)
While this procedure provides the optimal EVP solution for a system of intersections, it was found that when approaching the optimal solution, a very small improvement in ERV
191

travel time could require significant increases in the amount of preemption, and thus much higher side side-street delays. It was observed, in certain instances, that to achieve a gain on the order of 0.1 seconds in ERV response time through an intersection, preempt could increase on the order of 100 seconds. To avoid such instances, a more relaxed optimization criterion is placed where the search is limited to an ERV intersection-level travel time within 2-seconds of the lowest achievable travel time. Hence, a modification to the existing algorithm (in Figure 46) will be reflected in the flowchart in Figure 48.
Figure 48. Matrix. Flowchart for `Relaxed' Optimal EVP Call using VISSIM simulation iterations.
192

Development of Training Data Figure 49 shows the entire ML development process from creation of the input matrix to completion of the ML-model ready for recommending timing of EVP at an intersection.
193

Figure 49. Matrix. Flowchart for ML-model Development and Implementation
The object of the ML-model is to determine at each time step (i.e., each second as the ERV approaches the intersection) if preemption should be activated. Rather than employing queue length, a parameter that would require CV penetration rates or detection equipment unlikely in the foreseeable future, traffic parameters that are realistically extractable in real-time are used as features in the ML-model. This process also has the advantage of increased transferrable to other networks as the data leveraged is common across signalized corridors.
Input Matrix formulation
There is a need to list the features that could be extracted in real-time from a network. The ML algorithm uses these features to drive the decision for EVP call time at a signalized intersection. It is assumed that time vs location information is available for the ERV (which is the case with our current PIB network). Additional data at any time t that can be used are detector loop data (advanced detection, installed in the range of 200-500 ft upstream of the stop-bar in the mainline PIB), signal state information, and preemption state information of immediate upstream intersections in the preceding few seconds.
In addition, data used in this effort includes time elapsed since the ERV entered the corridor (i.e., in the PIB model this is the time the vehicle entered the network, in a field scenario this could be the time the ERV turns onto the corridor). These quantities are used together to create the training data input matrix as building blocks for the MLmodel. Figure 50 shows how the quantities can be placed in vectors as input to train the
194

signal to decide if an EVP call needs to be made (output=1) or NOT (output=1) at a given time.
Figure 50. Multiple elements. Schematic Representation of Input Vector of a PIB intersection
The input matrix for an intersection is filled from the time when the ERV enters the network to the time an EVP call is made at the intersection, if an EVP call needs to be made. If in a scenario an EVP call is not needed, the input matrix uses the data from the time of the ERV entry until the ERV reaches a minimum safe distance to the intersection (a point beyond which all emergency scenarios will get a mandatory EVP call for all intersections). Using Figure 51as reference, the preempt output will be 0 from the time the ERV enters the network and it becomes 1 at the time-stamp when the EVP call is made. Hence, the aim of the model is to predict the occurrence of output=1 among the 0 outputs. Once the EVP call is made, the call stays active at the intersection until the ERV crosses the stop-bar. Hence, the problem ends at the intersection after the call is made in case an EVP call needs to be made.
195

Figure 51. Matrix. Schematic Representation of Input Matrix and corresponding EVP output for a PIB intersection
The EVP problem is solved at a system level (i.e., for a designated ERV route) as a combination of the solutions at the individual intersection level. An EVP model is created for each individual intersection, and once the ERV indicates the route it wants to follow, the intersections comprising that route trigger their individual intersection level EVP recommender models concurrently to generate the EVP trigger times at each intersection. However, a drawback of this approach (i.e., ERV route-level) is that any failure at the sub-system level (such as gridlock) would not be incorporated in the model. To account for such a scenario, and to ensure safe passage of ERV, an EVP call is made at the predetermined "minimum failsafe" distance of the ERV from the signal, in the case that the ML model does not make the EVP call before the ERV reaches that failsafe distance. Hence, for a given scenario, the training matrix is only created either until the ERV reaches that minimum failsafe distance or until an EVP call is indicated per the "relaxed" optimal solution recommendation.
Splitting Dataset into Train-Test-Validation Components
196

Before implementing the ML-model on the created matrix, an essential step is to split the data into training, validation, and testing data. The rationale behind an appropriate split is important and there are several methods for determining the split. Of the 160 overall scenarios, 96 scenarios were chosen for training, 32 scenarios were chosen for validation and remaining 32 scenarios were chosen for testing (thus splitting the data as 60% for training, 20% for testing and 20% for validation), a typical ratio. Several 60-20-20 splits were made at random and the final split was chosen from among them ensuring that each subset had a mix of early EVP and late EVP for all intersections concerned while also ensuring that each subset had ample representation with respect to random simulation seed and ERV entry-time to the network.
Issues with the Classification Problem and the need for a Soft-Labeling Solution To perform a classification prediction model on the current input matrix, the task is to find the correct `1' among hundreds of `0's. In other words, less than one percent of output values are 1s. Hence, if a model predicts a 0 for all observations, it has a prediction rate higher than 99 percent, indicating a `good' model in the traditional sense. However, such an outcome essentially never predicts an EVP call, resulting in a useless outcome in the current context.
Additionally, as the wall-clock time (i.e., real-time) nears the optimal pre-empt call time the additional impact of delay to the conflicting movements due to an early EVP call lessens, that is, if the EVP occurs two second before the optimal time the impact on the conflicting movements is less than if the impact call occurs 20 seconds prior to the
197

optimal time. Hence, rather than an abrupt jump from 0 to 1 at the time of EVP, a method is needed to show that the need for EVP increases at an intersection with the passage of time. In other words, there needs to be a method to show that the input information closer to the optimal EVP call is more important for learning the model's EVP propensity. In such cases, a "soft-labeling" approach (Nguyen et al., 2014) was adopted that labelled the 0s (from ERV entry into the network) until 1 (at EVP call), monotonically increasing (indicating progressively increasing need for EVP with time) to reach a pre-determined cut-off. Passing the cut-off would imply an immediate need for EVP.
The soft-labelled value, in essence, would be a measure of the need of EVP at a given point in time for the intersection, which may be referred to as an "EVP Necessity Score" or EVPNS. It has been observed that the need for EVP varies with time at different intersections. Since there is no known literature on the nature of the EVPNS curve with time, several forms of monotonically increasing curves were chosen for testing; either convex, concave, or linear with a mix of exponential, sigmoid, negative-exponential, and linear functions of time to represent the curve. The upcoming steps describe the process used to decide what curve fits best in each situation. The schematic in Figure 52 explains this phenomenon. The choice of the final EVPNS curve and the corresponding cut-off was made based on how the ML-model performed based on this soft labelling of the validation set.
198

Figure 52. Graph. Demonstration of Soft-labeling decision-making
Choice of the best model The problem to be solved consists of two simultaneous sub-problems; firstly, the need to identify situations where no EVP is needed. That can happen where the ERV enters the upstream approach at a time such that the vehicle has the capability to cross the intersection on a GREEN indication without preemption. Secondly, if EVP is needed, the EVP prediction model has to identify when the call needs to be made. Hence, there are three parameters to control:
(4) No Preempt Threshold (NoPrThres): In cases where no preemption is needed, the EVPNS curve will go from 0 to this value (which is less than 1). Hence having a cap lower than one naturally helps the model to learn cases where the EVPNS does not reach the cutoff, thus, not triggering an EVP call. In cases that need EVP, the NoPrThres parameter has a value 1 by default. 199

(5) The form of the EVPNS curve: The form of the curve ranging from 0 up to 1 (where EVP is needed) or 0 up to NoPrThres (where EVP is not needed).
(6) The cutoff: The value that determines if an EVP will be called. When the EVPNS goes beyond the cutoff, an EVP call will be made. Figure 53 demonstrates the functionalities of the three soft-labeling parameters described here. Each line represents a different potential for the curve.
(a)
200

(b)
Figure 53. Graphs. Demonstration of use of Soft-Labeling Parameters: (a) Case 1: EVP call is needed; (b) Case 1: EVP Call is not needed and NoPrThres<1 is used.
Training the model
Multiple combinations of NoPrThres, EVPNS curve, and cutoff were used to make the necessary modifications to the training and validation matrices. Supervised-learning was performed to solve this regression problem. Multiple Machine Learning Algorithms were chosen and tested, namely, (1) Support Vector Regressor (SVR), (2) Neural Network and (3) Random Forest Regressor. For all the cases, k-fold cross validation was performed to choose the best performing hyper-parameters and mean-squared error was used as the loss function, in attempt to reduce the propensity to overfit or have (Shulga, 2018).
There are cases where there were very few data-points to train on, for example, cases where EVP is triggered soon after the ERV entry into the network, creating a few preEVP data-points in those scenarios. Hence, cross-validation was preferred over hold-out to allow the model to train for various train-test splits, making it more well-informed to make calls on unseen data-sets (Allibhai, 2018). For all ML-procedures, Python's Scikitlearn Package was used (Scikit-learn). The following is the list of hyper-parameters that were tuned for each of the ML algorithms:
(1) Support Vector Regression: The problem to be solved at the intersection level can be described as determining the ideal preemption time point to allow the queues to clear such that the ERV experience minimize hindrance while the side street of conflicting movements do not experience undue delay. With the combination of acceleration and deceleration of multiple vehicles, as well as the
201

continued movement of non-queued vehicles required to achieve this outcome, the underlying problem is likely to be non-linear. Support Vector Regression (SVR) uses the same principle as the SVMs. The basic idea behind SVR is to find the best fit line. In SVR, the best fit line is the hyperplane that has the maximum number of points. Hence, an SVR that acknowledges the non-linearity in the data while providing a proficient model is considered (Raj, 2020) .To train with SVR, the following three hyper-parameters were tuned in the model:
(a) Kernel: `linear', `poly' (polynomial), `rbf' (radial basis function), `sigmoid'
(b) Degree: Considered degrees 2-5 (only for `poly' kernel)
(c) gamma: Kernel coefficient for `rbf', `poly', and `sigmoid' kernels. It controls the sensitivity to differences in feature vectors.
(2) Neural Network Regressor (Scikit-learn, #56)
A problem such as this is expected to give an outcome as a function of certain variables. Most real-world problems such as these can have an approximated solution built using Neural Networks (Fortuner, 2017 #134). This problem consists of a few traffic parameters that are expected to contribute to the eventual EVP decision. The dependencies of the decision on the parameters considered is usually unknown. Since the objective of this task was to develop a "black box" decision maker that could produce an approximate solution for this problem, use of a Neural Network Regressor was explored. Several hyper-parameters were tuned for this algorithm, keeping the `solver' algorithm fixed at `adam' for optimization. The Adam optimization algorithm is an extension to the stochastic gradient descent approach and is more suited for our medium-sized data set (>100 MB for all intersections in question). This approach has been documented to give several benefits, such as straightforwardness in implementation, computational-
202

efficiency, low memory requirement, and well-suited for non-stationary objectives (Brownlee, 2017 #135).The following hyper-parameters were tuned for build out of the Neural Network model:
(a) `activation': Decides the activation function for the hidden layer. It could range from identity to several linear and non-linear dependencies (Sharma S. (2017)). The options in scikit-learn are `identity', `logistic' (logistic sigmoid function, i.e., f(x) = 1 / (1 + exp(-x)), `tanh' (hyperbolic tangent, i.e., f(x) = tanh(x)) or `relu' (rectified linear unit function, i.e., f(x) = max(0, x)).
(b) `hidden_layer_sizes': This is an array that defines the number of hidden layers and the number of neurons in each layer. Several combinations were explored in the cross-validation stage.
(3) Random Forest Regressor (Scikit-learn, #57)
While SVR and the Neural Net Regressor are expected to handle problems with an underlying complicated prediction model, it is essential to test a relatively simpler algorithm. If Random Forest Regressor provides comparable results, the upside is significant for two reasons, (a) the training time would be significantly lower, and (b) the underlying model would be significantly simpler providing possibilities for easier transferability of concept into similar problems. This algorithm uses the ensemble learning method for regression, where it combines multiple regression algorithms to get a more accurate algorithm (Bakshi, 2020 #136). Two hyper-parameters were tuned for this algorithm:
(a) `max_depth': This controls the max_depth to which the tree will be allowed to reach with the random forest structure.
203

(b) `max_features': If the total_number of features is given by n_features, this hyper-parameter decides as a function of n_features, how many features would be used: `sqrt' (square root of n_features), `log2' (log base 2 of n_features), `auto' (max_features = n_features), etc.
Evaluating and choosing the best prediction model
For each intersection, multiple trained models were chosen, combing soft-labeling curve (linear, convex exponential, concave exponential, quadratic), ML-algorithm, and Cutoff and NoPrThres. These models were trained on the training data with the process previously described and then were tested on the validation data-set.
Choosing a model as the `best' model for a given intersection requires a definition for the errors that we are trying to minimize. For this task we considered two conditions: (1) the duration for which EVP call is made, but was NOT needed (i.e., a False Positive) and (2) duration for which EVP call was needed, but the model did NOT activate EVP (i.e., a False Negative). The first determines how much extra EVP is given by the model in question that, in turn, will increase delay on the conflicting approaches, the reduction of which is a secondary objective of the project. The second indicates a missed/delayed EVP call, which can cause additional delay in ERV response, the reduction of which is the primary objective of the model.
Hence, to be consistent with the primary objective of the study, false negatives are weighed ten times more than the false positives. This is a decision made based on how much priority a strategy seeks to give to having excess preempt time at the cost of delay
204

to the side-street traffic. This factor could vary based on the general demand in the conflicting movement, time of the day, and general history of what routes are more in demand in emergency conditions. An additional factor in this experiment that plays a role is a True Positive (i.e., an EVP call was needed and was made correctly). The mixed weighted number created with the False Negatives and the False Positives is normalized by the number of true positives. Hence, the best model is defined by a variably weighted factor defined as follows:

. - + -

- - =

-

()

The optimization goal is to have the lowest score possible. In simpler terms, the factor is expected to let one know "how much does a given ML-model get wrong out of the number of seconds the model gets the EVP call right". With that score in place, the top few model-combinations giving the lowest EVP-model-score were chosen and used to run the scenarios in the validation data, as another level of the selection process. Using the PIB@ North Berkeley Lake Road as example, Table 6 shows us first few rows of how various combinations work to give the lowest (best) EVP-model-scores for the EVP MLmodel training at that intersection. From the top of the list, three model-combinations are chosen and VISSIM simulations are run on the validation set to choose the `best' among them.

The three models chosen based on their low EVP-model-score were used for further
experiments had the common features: Soft_label: Negative Exponential; ML_algorithm:
Neural Network, NoPrThres: 0.9 and they varied in the cutoff feature: (1) cutoff=0.9, (2) 205

cutoff=0.91, (3) cutoff= 0.92. As seen in Figure 54(a), very little difference is found between the intersection level travel-time for the three chosen models. Figure 54(b) shows that there is negligible difference in the overall intersection-level preempt duration as well. So, the model with the lowest travel-time, i.e., the model with cutoff=0.9, was chosen (even though the difference is not significant). It is important to note that this final level of model choice is not expected to have significant effect in the overall results, but it is still performed to have a process elimination in the case of a rare scenario where a `good' model gives unexpected or unrealistic result when tested on a real-simulation model.
Table 6. First columns of EVP model results for Validation data (PIB @ North Berkeley Lake Road)
206

(a)
(b) 207

Figure 54. Graphs. Validation Set ERV KPI Comparison with Different EVP-ML-models: (a) Travel-time, (b) Preempt Duration
ML-model Discussion The parameters for the best EVP recommender model for each intersection is summarized in Table 7. In all cases, the Neural Net Regressor model was identified as the best model. Support Vector Regression did not yield encouraging results for any of the intersections. Simple Logistic Regression yielded results comparable to the Neural Net Regressors approach only for the North Berkeley Lake Road intersection and were suboptimal elsewhere.
One of the key drawbacks of Support Vector Machines is that the algorithm typically fails to work for (1) large datasets, (2) datasets where number of features exceeds the number of training data samples, or (3) noisy datasets (Raj, A. (2020)). The first two of these issues being a significant issue for these data. The training matrix uses a large number of features, since it uses information from the last 160 seconds as individual features at any given simulation second. Hence, at least at a scenario level, the data is likely insufficient to fully train an SVR implementation. If the training data were significantly expanded (e.g., use of 10-20 random seeds to generate a larger volume of training data) than exploring SVR again could be a direction to explore. Also, it is worth noting that the training time for SVR was also significantly higher than the other algorithms.
Table 7. Summary of the final intersection level ML-models for the seven contiguous intersections involved in this study
208

Intersection

NoPrThres

South Old Peachtree Rd 0.95

Highwoods Center Pickneyville Park

0.95 0.925

South Berkeley Lake Rd 0.93

Soft-labeling function

Cutoff

Positive Exponential 0.94

Linear

0.95

Negative Exponential 0.925

Linear

0.93

North Berkeley Lake Rd 0.9

Negative Exponential 0.9

Summit Ridge Rd

0.98

Sigmoid

0.9

Howell Ferry Rd

0.9

Positive Exponential 0.9

As illustrated in Figure 55, all the intersections on the ERV route, excluding PIB@ North Berkeley Lake Road), were sufficiently close (less than or equal to 3000 feet) to interact with its upstream intersection. The interaction between an intersection and its upstream counterpart makes the underlying problem more complicated than in a simple linear model, especially when they are in close proximity. While Logistic Regression worked comparatively well at PIB @ North Berkeley lake Road the Neural Network Approach still performed better. The Neural Network Approach is thus the preferred approach for all of the intersections. The sample code and the steps followed in running them from training to validation are available at the GitHub repository at:

209

https://github.com/hunter-guin-gatech/Peachtree-industrial-Boulevard-ML-modelDevelopment
Figure 55. Photo. PIB corridor indicating PIB @ North Berkeley Lake Road being placed significantly downstream (~6500 ft) of PIB @ South Berkeley Lake Road
Evaluation of the ML-Model In the previous sections, the ML strategy had been discussed at the individual intersection level. The following discussions will focus on the integration of these individual component ML models into an overall ML system model and the testing of this model at the network level. As for the DP algorithm, the individual ML-models were run on the seven intersections concurrently with the ERV placing a EVP call to the PIB @ Medlock
210

Bridge intersection upon entry into the network with the remaining intersections having their EVP called based on the ML algorithm. As before, a 60-20-20 train-validation-test split was performed with 32 of the 160 scenarios set aside for the test-case with the remaining used for the training and validation activities. Model sensitivity experiments were performed on these test data and focused on: (1) Changes in driving factors of the ML-model, viz. random seed, ERV arrival time, etc., and (2) Adjustments in the experimental strategy, such as multiple ERVs in close succession, testing it with SILS architecture, etc. Summaries of the results from these tests are discussed in subsequent sections. Model Performance The overall network-level performance of the DP and ML algorithms are compared with the "relaxed optimal" (i.e., perfect knowledge) solution . Figure 56(a) presents the results in terms of ERV speed while Figure 56(b) presents the results in terms of ERV travel time.
211

(a)
(b)
Figure 56. Graphs. Comparative Results in Violin-plots: ML-Prediction model vs other EVP call experiments on the test-data: Mainline-PIB: South Old Peachtree Rd to Howell Ferry Rd: (a) ERV
Speed (mph), (b) ERV Travel Time (s)
212

In Figure 56(a), the ML-model, represented by the yellow violin plot, aligns well with the grey violin plot, which represents the `relaxed' optimal solution. On a 45-mph road, in a PM peak environment, the ML-model gives, on an average, a 42.5 mph speed through the corridor, which is within 10% of the posted speed limit. The DP solution (i.e., the 4n+9 solution), represented in pink, provides a slightly less desirable solution (<40 mph).
Similar interpretations can be made for Figure 56(b). While the ML-model (yellow) does not provide the optimal travel time as compared to the absolute global optimal solution (represented by the dark green violin plot), it does provide comparable overall travel time compared to the ideal solution (in grey), which is a `relaxed' optimal solution. Based on a KW-test, it was concluded that that the ERV travel-time in the `relaxed' optimal solution (in grey) and the ML-model driven solution (in yellow) are not statistically different.
Another aspect that needs to be considered is the length of the preempt call placed in each strategy. A longer EVP call can produce good performance for the ERV but be suboptimal for the conflicting routes. Figure B15 is a violin-plot representation of the total preempt duration for the seven intersections subjected to the ML-model (yellow), when compared to the `relaxed' optimal solution (grey), `extreme' optimal solution (dark green) and the DP solution (pink). Applying the KW-test here, it is found that the preempt duration with the ML-model and the one given by the `relaxed' optimal solution, do not have a statistically significant difference. This provides initial evidence that a trained ML-model, working with limited data input, can perform as well as more dataintensive strategies.
213

Figure 57. Graph. Aggregate Preempt Duration on seven interchanges at PIB for Different Experimental Strategies
The mainline and side-street travel-times for non-ERV vehicles for the different strategies are shown in Figure B16. Figure 57(b) indicates that the DP solution has more side-street travel disruption compared to the ML-EVP model with the ML-EVP model (yellow) travel-time for non-ERVs adheres closely to the corresponding travel-time for ERVs from the "ideal" solution for mean and median travel times. This difference is statistically significant based on a KW-test. This provides another strong argument for using the EVP-ML-model in lieu of heuristic DP model. Figure 58(a) also indicates an improvement in travel-time for the non-ERVs that share the same trajectory as an ERV however these differences are not statistically different between the DP and ML models.
214

(a)

(b)

Figure 58. Graphs. Non-ERV Travel time Variation with respect to ERV trajectory or Preemption Event: (a) Along main-line PIB, (b) All side-street vehicles since an EVP event Model Sensitivity Analysis
The ML-model was built using a microscopic simulation model that was calibrated to reflect certain traffic flow and ERV behaviors that were observed in the field. While this process gives a certain degree of validity to the results produced by the simulation model, there are assumptions regarding the behavior patterns of vehicles in the simulation model, which might be violated in the real world under certain circumstances. For example, the simulated ERV never crosses the centerline, or goes over the shoulder, while real world ERVs do perform such maneuvers.

Another issue that can arise is overfitting (use of too many variables) which makes the trained model too specific to the training conditions. Sensitivity analysis can aid in
215

determining the presence of overfitting by analyzing the sensitivity of the model outputs to changes in the underlying variables. Several aspects of the model can be adjusted for this type of analysis with the choice of the parameters guided by the variables that played a significant role in model training. The results of these analyses are described below.
Sensitivity Analysis I: Varying random seed
The random seed value in a VISSIM simulation initializes the random number generator controlling the stochastic variations in various traffic parameters, such as vehicle arrival, speed, acceleration, travel volume, etc. Two simulation runs with identical input files will show difference in those traffic characteristics when run with different random seeds. Hence, varying the random seeds would test the response of EVP ML-model to slight stochastic variations in the traffic conditions that could partially replicate the variations that are observed on different days.
The training, validation, and test sets are all built with variations of simulation runs with random seeds 1 to 5. Then, for the sensitivity analysis, the effectiveness of the model is tested with random seeds 6 to 10, keeping everything else identical. Figure 59 illustrates how the ERV speed within the various model strategies with Figure Figure 59(a) showing the previous results (random seeds 1-5) and Figure 59(b) showing the results for the new random seeds (random seeds 6-10). The ML results for the random seeds 6-10 gave an average ERV speed of 40.1 mph that is somewhat lower than the 42.5 mph average speed for random seeds 1 to 5. However, similar results are observed in the other strategies as well indicating that it is a system-level rather than a strategy-level difference.
216

When travel time is used as the performance measure Figure 59(b), on average the MLmodel has an overall improvement of 160.6 seconds in travel time with random seeds 6 to 10 compared with the 122 s improvement in the original strategy involving seeds 1 to 5 with little, if any, difference in the relative performance of the various strategies.
(a)
(b) 217

Figure 59. Graphs. Comparison of ERV KPIs with Changed Simulation Random Seeds: (a) Speed [mph], (b) Travel time [s]
Sensitivity Analysis II: Scaling down Traffic Volume The simulation model used for testing was based on the PM peak hour. Since the models were tested (and trained) in a high congestion scenario, there is a possibility of obtaining different results if the network volumes were lower. Hence, this condition was tested by running the simulation with the network wide volumes scaled down by 0.5X and 0.75X in two separate experiments, keeping all other elements identical (e.g., no changes in O-D patterns or signal timing). Figure 60 shows the resulting ERV speeds and travel times for the baseline and reduced volume scenarios. As before, the ML-model (in dark green) performed consistently perform better than the DP algorithm (in pink). As expected, higher ERV speed was achieved due to volume being reduced along the corridor.
218

(a)
(b)
Figure 60. Graphs. Comparison of ERV KPIs with Scaled-down Volume: (a) Speed [mph], (b) Travel time [s]
Sensitivity Analysis III: Varying ERV Entry Time into the PIB network One of the primary inputs in the simulation model is the time that the ERV is introduced into the network. In the baseline model the ERV was introduced at approximately the one-hour mark of the simulation (i.e. approximately 5:30 PM). To test the sensitivity of models to this factor in the resulting models, the time the ERV was introduced into the network by 800 seconds (i.e., five signal cycle lengths) earlier and later than for the baseline runs keeping everything else in the modeling constant. scenarios identical. Results of these tests are shown in Figure 61 for ERV speed and travel-time. These
219

results remain consistent across all strategies with the ML-model giving travel-time improvements in a similar range as the original strategy and maintaining an average ERV speed above 40 mph.

(a)

(b)

Figure 61. Graphs. Comparison of ERV KPIs with Varying ERV Entry time into the Network: (a) Speed [mph], (b) Travel time [s]
Sensitivity Analysis IV: Introducing two ERV vehicles in tandem to the same destination.
Situations can arise where more than one ERV approaches an intersection. They could be arriving on the same approach or on conflicting approaches at the intersection. This experiment focuses on the case where two ERVs are approaching the intersection from the same direction and tests the efficiency of the model under the concurrent EVP call scenario. Figure 62 shows trajectories of ERVs in tandem with two different levels of gaps between them of 10 and 800 seconds.
220

(a)
(b)
Figure 62. Graphs. Trajectory of vehicles around and including ERVs in tandem with varying headways between the two ERVs (represented by blue and purple lines): (a) 10s gap (b) 800s gap.
Figure B21 shows how the ERVs in the tandem scenario affect the average speed of the leading and following vehicles in 16 different cases, with gaps ranging between 10
221

seconds to 960 seconds. As general intuition would dictate, the leading ERV will not be affected by the current strategy and the model's effectiveness will primarily be tested by how the EVP calls clear out traffic for the following ERV.
Figure 63. Graph. Speed variation of Leading and Following Vehicles for ERVs in Tandem Experiment with varying gap between the two ERVs.
In Figure 63 ERVs with gaps between them of 90 seconds or less show no meaningful differences between the speed profiles of the leading and following ERVs as both are served with a single `long' EVP call at all/most intersections. Focusing on the gaps highlighted by the orange box, the speed of the following ERV begins to be reduced reaching a minimum with about a 200s gap before beginning to recover. The most likely explanation is that the following ERV only makes its EVP call after the call of the leading vehicle had been completed thereby introducing delay associated with the signal timing transition that occurs between two EVP calls. For example, after completion of the first call the signal controller may need to wait for the service of the minimum green of
222

the current call. At longer time intervals, the effect diminishes until the two ERVs calls can be viewed as independent.
Sensitivity Analysis V: Varying Entry-point of ERV One of the factors in the input matrix for ML-model training is the cumulative distance covered by the ERV in the network. This variable was designed to establish how far away the ERV is from a given interchange based on a consistent entry point. To examine the generalizability of this parameter, the original route was split into two partial routes for ERVs to create two separate experiments: Path 1: Medlock Bridge Road to Pickneyville Park, Experiment 2: South Berkeley Lake Road to Howell Ferry Road. Figure 64indicates the full route and paths 1 and 2, respectively.
(a)
223

(b)

(c)

Figure 64. Photos. ERV Routes: Varying Experimental Strategy on mainline PIB: (a) Original Route: Medlock Bridge Road to Howell Ferry Road, (b) Path 1: Medlock Bridge Road to Pickneyville Park, (c) Path 2: South Berkeley Lake Road to Howell Ferry Road.

Since the ML model was trained to function for the original ERV route translational modifications need to be made for distances and times for the ML-models of the partial paths. If the new start point is distance d2 upstream from an intersection and the original ERV entry-point was distance d1 upstream, then the intersection distance must be translated forward by (d1 d2). A similar translation is needed for the time parameter T. Assuming an expected speed of v for an ERV T needs to be translated forward by (d1 d2)/v. Note that depending on the time of the day or general requirements of the network, v could vary. For these tests, v =40 mph was used.

Figure 65 shows the speed results from the translated ML-models. In both cases, the translated models achieved average ERV speeds above 40 mph, which is better than that achieved with the DP strategy. Figure 66 shows the same results in terms of travel time.
224

The ML-model results in an average ERV response time 127.3 seconds better than the no EVP condition over a total of distance of approximately 4.7 miles.

(a)

(b)

Figure 65. Graphs. Comparison of ERV Speed with Modified ERV Routes: (a) Path 1: Medlock Bridge Road to Pickneyville Park, (b) Path 2: South Berkeley Lake Road to Howell Ferry Road.

(a)

(b)

Figure 66. Graphs. Comparison of ERV Travel Time with Modified ERV Routes: (a) Path 1: Medlock Bridge Road to Pickneyville Park, (b) Path 2: South Berkeley Lake Road to Howell Ferry
Road.

225

Model Testing in a SILS Environment Since the EVP ML-models were trained using VISSIM's RBC as the signal controllers, the models were also tested using software emulations of actual field controllers used at intersections throughout the State of Georgia. This SILS architecture was built using VISSIM and MaxTime and is the same configuration used to test the DP method. The MaxTime signal controller software also features a different exit strategy referred to as the "Queue Recovery" exit-transition. "Queue Recovery" keeps track of the EVP calls and recalls in the detector and serves the longest waiting phases. This feature was designed to offer an effective way of handling conflicting movements in the aftermath of an EVP call. The performance of the "Queue Recovery" transition was included in the tests along with the default exit-transition strategy.
Figure 67(a) presents the ML-model results with SILS as well as the corresponding results with the internally programmed VISSIM RBCs. On average, the predicted speeds with the SILS, while similar to the results with VISSIM RBCs, were consistently 1-2 mph lower for all the tested strategies when compared to their RBC counterparts. However, as shown in Figure 67(b) the predicted improvement for ERV travel times for the ML-model over "No Preempt" cases was approximately 115 seconds for both the SILS and RBC architectures. Hence, it appears that the ML-model strategy is not affected any differently than the other EVP strategies, or no-preemption case, by changing from the RBC architecture to the SILS architecture.
226

(a)

(b)

Figure 67. Graphs. Comparison of ERV KPIs under SILS architecture compared to RBC model: (a) Speed, (b) Travel times.

Another aspect that was investigated was the effect of using Queue Recovery exittransition. A Queue Recovery exit-transition prioritizes the movement that has more queues resulting from the EVP call and serves the corresponding phase. Figure 68 shows how different EVP methods cause varying travel times on the side-streets when EB and WB movements at PIB are aggregated for all seven EVP-programmed intersections going from South Old Peachtree Road to Howell Ferry Road. The figure shows the distributions of side-street travel times and the progression of those with passage of time.

As we see, the DP solution and ML-Predicted solution (blue and orange respectively) non-ERV travel-times under default exit-transition settings are consistently higher than the Normal exit RBC setting (light green violin-plot). The effect is reduced as time exceeds two signal cycle lengths (2*160s). With the Queue-recovery setting of SILS (cyan violin-plot), the side-street non-ERV travel times are significantly lower and is somewhat lower than the RBC driven scenario.
227

In contrast ERV speeds does not receive any meaningful adverse effect from the Queuerecovery exit transition (green violin plot: average speed: 39.4 mph) when compared to the default setting SILS case (turquoise violin-plot: average speed: 39.7 mph).
Figure 68. Graph. Comparison of non-ERV Side-street travel-time after an EVP Call: Variations of EVP call strategy under SILS Architecture
While the difference is negligible even under the current strategy, the effect on any difference from this remains to be studied if actual vehicle pull-over behavior in response of ERV is simulated within the model. ML-model Runtime discussion The experiments in the previous section showed that the ML-model functions in a robust fashion under various scenarios and shows promise for transferability of the methodology into other networks and different traffic flow conditions. It is important, however, to also
228

ensure that the model can be incorporated in real-time field deployment setups. Hence it is important to test the computational efficiency of the model and ensure the feasibility of running the model faster than real-time so that the decisions can be obtained in time to act.
To estimate the run-time efficiency, experiments were performed for both the RBCdriven and SILS models. The simulation models were run with and without the MLmodel integration to obtain an estimate of the run time execution impact of the MLmodel. In the absence of the ML-model, the EVP is run by the simply triggering EVP at the same times that the ML-model was observed to trigger EVP, all other conditions being held equal.
Figure 69 shows how the different experimental setups perform when their simulation clocks are compared to the wall-clock over a period of about 5 minutes of simulation time when the ML-model is active in the simulation. The same 5-minutes of simulation time are extracted from the without-ML-model runs for comparison. The darkened line on the plot shows the average trendline for simulation time versus wall-clock time over 10 replicate runs with 10 different random seeds. The shades around the dark lines show minimum and maximum variation in simulation speed at different instances.
All runs are made on a workstation computer with an Intel Core TM i7-7700 CPU @ 3.60GHz and 16GB of RAM. All trend-lines were approximately linear for the entire EVP call-period, on average. For the SILS model runs, the simulation typically runs at 5X real-time speed without the ML-model and runs 4X real-time speed with the ML-
229

model. For RBC-driven VISSIM simulation, the run works at 12X real-time speed without the ML-model and at 8X real-time speed with the ML-model. In an actual deployment, the computer will have to run just the ML-model. Hence the load on the computer will be even lower than in the current experiment. However, this experiment gives a relative measure of the impact of the ML-model under different computational loads for the different simulation models and provides sufficient evidence that the MLmodel can perform faster than real time.
Figure 69. Graph. Comparison of VISSIM simulation Runtime, in comparison to wall-clock time, with both VISSIM-RBC and SILS architectures
REFERENCES for Appendix B Allibhai, J. (2018). Hold-out vs. Cross-validation in Machine Learning.
https://medium.com/@jaz1/holdout-vs-cross-validation-in-machine-learning7637112d3f8f#:~:text=Cross%2Dvalidation%20is%20usually%20the,just%20one %20train%2Dtest%20split Chang, E., & Messer, C. J. (1991). Arterial signal timing optimization using PASSER II90-Program user's manual.
230

Chen, Z., & Huang, X. (2017). End-to-end learning for lane keeping of self-driving cars. 2017 IEEE Intelligent Vehicles Symposium (IV),
Cohen, S. (1982). Concurrent use of the MAXBAND and TRANSYT signal timing programs for arterial signal optimization. Federal Highway Administration, Urban Transportation Management Division.
Cong Minh Ho Chi, C., Matsumoto, S., Sano, K., Cong Minh Doctoral Student, C., Matsumoto Professor, S., & Sano Associate Professor, K. (2005). The speed, flow and headway analyses of motorcycle traffic Title: Two-player Game Theory Based Analysis of Motorcycle Driver's Behavior At Signalized Intersection View project THE SPEED, FLOW AND HEADWAY ANALYSES OF MOTORCYCLE TRAFFIC. https://doi.org/10.11175/easts.6.1496
Grzywaczewski, A. (2017). Training ai for self-driving vehicles: the challenge of scale. Available from Internet: https://devblogs. nvidia. com/training-self-drivingvehicles-challenge-scale.
Hse, F., Roch, L. M., & Aspuru-Guzik, A. (2019). Next-generation experimentation with self-driving laboratories. Trends in Chemistry, 1(3), 282-291.
Hossain, S., Fayjie, A. R., Doukhi, O., & Lee, D.-j. (2018). CAIAS simulator: selfdriving vehicle simulator for AI research. International Conference on Intelligent Computing & Optimization,
Kamble, S., & Kounte, M. R. (2022). A Survey on Emergency Vehicle Preemption Methods Based on Routing and Scheduling. International Journal of Computer Networks And Applications, 9, 60-71. https://doi.org/10.22247/ijcna/2022/211623
Min, W., Yu, L., Chen, P., Zhang, M., Liu, Y., & Wang, J. (2019). On-demand greenwave for emergency vehicles in a time-varying road network with uncertainties. IEEE Transactions on Intelligent Transportation Systems, 21(7), 3056-3068.
Mohammadi, R., Roncoli, C., & Mladenovic, M. N. (2020, 3-5 Nov. 2020). Transit signal priority in a connected vehicle environment: User throughput and schedule delay optimization approach. 2020 Forum on Integrated and Sustainable Transportation Systems (FISTS),
Nguyen, Q., Valizadegan, H., & Hauskrecht, M. (2014). Learning classification models with soft-label information. Journal of the American Medical Informatics Association, 21(3), 501-508.
231

Raj, A. (2020). Unlocking the True Power of Support Vector Regression. Retrieved March 15 from https://towardsdatascience.com/unlocking-the-true-power-ofsupport-vector-regression-847fd123a4a0
Rao, Q., & Frtunikj, J. (2018). Deep learning for self-driving cars: Chances and challenges. Proceedings of the 1st International Workshop on Software Engineering for AI in Autonomous Systems,
Scikit-learn. scikit-learn: sklearn.neural_network.MLPRegressor. Retrieved July 15 from https://scikitlearn.org/stable/modules/generated/sklearn.neural_network.MLPRegressor.html
Scikit-learn. sklearn.ensemble.RandomForestRegressor. Retrieved September 12 from https://scikitlearn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.ht ml
Shoufeng, L., Ximin, L., & Shiqiang, D. (2008). Revised MAXBAND model for bandwidth optimization of traffic flow dispersion. 2008 ISECS International Colloquium on Computing, Communication, Control, and Management,
Shulga, D. (2018). 5 Reasons why you should use Cross-Validation in your Data Science Projects. https://towardsdatascience.com/5-reasons-why-you-should-use-crossvalidation-in-your-data-science-project-8163311a1e79
232

APPENDIX C. VISSIM MODIFICATIONS TO INCORPORATE DRIVER PULLOVER BEHAVIOR AND COMPARISON OF EVP ACTUATION WITH AND WITHOUT DRIVER PULLOVER
Introduction Background Historically, studies and comparisons of preemption strategies have aimed to find the optimum trade-off in improving ERV travel times while minimizing the impact on nonERV travel times (How a Connected Emergency Vehicle Preemption System Works, 2019; Obrusnk et al., 2020; Shaaban et al., 2019b). Similar to these previous studies, the first two sections of the current study simulates the process through microsimulation of an ERV moving through an existing network and primarily focuses on the ERVs actions and route choices. This section of the study aims to improve the accuracy of the model by incorporating the microscopic interactions that occur between ERVs and non-ERVs in the real world due to the mandated requirement of non-ERVs to shift lanes to the right or pull-over to make way for an ERV. This section aims to evaluate the potential impacts of these vehicle-ERV interactions by modifying the default driver behavior in a microscopic simulation model to incorporate typical observed behavior of vehicles during encounters with ERVs, and to examine the effects of these behaviors on the results of preemption simulations.
233

The study focuses on two primary aspects of interaction. Unlike a normal vehicle, in congested conditions an ERV is expected to try to move over to the left-most lane to try to pass slower vehicles. The non-ERVs are mandated to move over to the right lanes and pull over safely when possible onto the shoulder to allow the emergency vehicle to use the left lane or left shoulder to overtake the non-ERVs. The lane-shifting and "pull over to the shoulder if needed" behaviors are jointly referred to as the pull-over behavior in this study and was is the primary focus of this study. There are several that can affect how drivers may react to the presence of an ERV on the road. Drivers can be expected to vary significantly in both their perceptions and reaction times to an ERV. Moreover, driver compliance with pull-over can vary as well. People may be unaware of the ERV until the vehicle comes within several feet, or they could be reluctant to give up their positions on the road for fear of being taken advantage of and overtaken by other nonemergency vehicles on the road.
Previous studies have acknowledged the importance of modeling these interactions and a few of them have studied the operational and safety impacts. In a recent study, Corts and Stefoni (2023) examined and simulated real-world behavior from video collected in Chile. The upstream and downstream distance ranges within which the ERVs are expected to influence a lane change for the non-ERVs were estimated from manual analysis of the video and was used to model the behavior of non-ERVs in the simulation model. It is however important to note that the number of data points on which the estimates were derived was quite small and limited in scope.
234

Implementations of ERV-traffic interactions have been developed on several simulation platforms. Zhang et al. (2009) modeled driving behaviors of non-ERVs alongside the ERV movement logic within CORSIMTM. The open-source SUMO platform was used by Weinert et al. (2019). Corts and Stefoni (2023) used PARAMICS.
This effort used PTV VISSIM and its External Driver Module (EDM) component to develop a realistic model of ERV/Driver interactions and examines the effects of preemption on ERV and general traffic travel times.
An extensive study conducted by Weinert and Dring (2015) surveyed 252 drivers of fire engines, ambulances, and police cars in Germany to better understand the risks involved during an emergency response. The survey provided strong evidence that non-ERV drivers have varying levels of reaction and response times to the presence of ERVs. Additional video data complemented their efforts to model a rescue lane approach in response to the presence of an ERV. Buchenscheit et al. (2009) also followed a similar methodology, in gauging the benefits of a proposed emergency vehicle warning system. However, the majority of these studies have investigated the impacts in the context of a V2X or a V2V communications (Buchenscheit et al., 2009; Lidestam et al., 2020; Savolainen et al., 2010). In comparison to a traditional system without communications between vehicles, the reactions of drivers in any system with V2X capabilities tend to be quicker. For the current section of the study, the focus was on simulating driver behavior without any assumptions about the existence of connected vehicle infrastructure. For the results presented here, the models assume that the ERVs have the capability to invoke a
235

preemption call either with V2I communication or with regular cellular communication between the ERV and the signal infrastructure. No other forms of communication is assumed between the vehicles exist within the simulation.
Data Sources Data Request To capture behavioral aspects of drivers, video data would provide the necessary breadth of information. However, video data was not readily available for use in this project. As an alternative, the research team sought input from people directly involved in and responsible for the operation of ERVs, including the county's Fire Chief. The responses from this focus group laid the foundations for the modeling assumptions. The primary questions posed to the focus group are listed below, and the conclusions from the responses are summarized in the next sub-section.
Questionnaire
When their sirens are on, do fire trucks always use the left lane or left shoulder? Does the emergency vehicle ever cross across the center of two lanes? Under what
conditions? How often (approximately) does this happen? Does the right shoulder ever need to be used? How often do you come across drivers who are unclear about what to do next? Is there a significant rate of pullover non-compliance? How about compliance when the ERV light is red at intersections?
236

How far in advance do drivers begin to shift lanes? In a similar vein, how long do drivers wait to return to their lane once the EV has
left? Does the presence of an EV cause the pace of other traffic to slow down? Do drivers typically leave just enough room for EVs to get through or do they often
make complete lane changes? What specific procedures do EV drivers take when they approach an intersection
with oncoming traffic on cross streets? Do the drivers on the cross streets always slow down to make room for it? Does pull over compliance differ significantly on single-lane roads versus multilane roads? What actions are recommended for other drivers to take while an EV is merging with their link?
Data Inference The following were the major conclusions from the responses from the focus group:
The ERV operators have a defined set of protocols which they must adhere to during emergency response. Despite this, most responses are very incident specific, and may call for non-standard procedures. An example of this is the use of non-paved areas on the road for traversal. Although such a maneuver is not recommended, it can be done if the circumstances deem it necessary.
237

The level of compliance for pull-overs is close to 100%. Most vehicles do follow the law and move a lane over, but the time taken to respond can vary greatly. There may be numerous reasons for such variability can be attributed to a multitude of factors including distraction and low levels of alertness, lack of available space for the drivers to move, etc. The ERV drivers have been trained to be patient and give the drivers time to execute their response.
There are situations where vehicles make a lane change to the left instead of the right, or pull over to the left shoulder, but these are usually observed on multi lane roads, near the intersections.
The pull-over responses can usually be seen up to 5-10 car lengths (100 to 200 ft) downstream. The responses of the ERV and the other vehicles are heavily influenced by road geometry, such as the presence of a median, the presence of a left shoulder, etc.
Lane partitioning or a "rescue lane" formation is usually observed in congested scenarios, where the ERV has no choice but to go through the middle of the traffic. However, ERVs try and use the leftmost lane as much as possible on their route to the emergency.
The general mandate for road users in Georgia are to yield the Right of Way by safely and quickly moving to the right-hand side of the road and come to a complete stop. If possible, they are to pull over to the right shoulder or the nearest available space to allow the emergency vehicle to pass. 238

o Modeling Components:
COM The component object model (COM) interface of VISSIM allows users to interact with the model via scripts. Any required information can be accessed and relayed during the runtime of a model. The scripts can be written in several languages and platforms (e.g., JavaTM, MATLAB, PythonTM etc.). For this study, the scripts were developed using PythonTM.
EDM External Driver Models (EDMs) are additional components that are used to extend the functionalities of VISSIM. EDMs are designed to enhance the current functionalities of VISSIM and give the user more control in "fine tuning" microscopic behavior. Two independent EDMs were designed to handle both behavior classes. These EDMs overwrite the existing defaults within VISSIM. The models are written in C++ and interfaced with VISSIM via Dynamically Linked Libraries (DLL). Throughout a simulation run, there is a constant exchange of data between VISSIM and the DLLs. The DLLs receive the vehicle data and return necessary instructions if defined conditions are met. The use of DLLs allows extensive control over car following models and lane change procedures. The general linkage and data exchange process between VISSIM and the EDM are visualized in Figure 70 below.
239

Figure 70. Matrix. Relationship between the EDM and VISSIM
Model Parameter Adjustments Several modifications and associated assumptions were required for modelling the ERV and non-ERV behavior to assimilate the results of the focus group study. The details are provided in the following subsections.
5) Veihcles impacted: Vehicles in VISSIM, by default, "see" two vehicles in front of and behind, and two lanes to either side (i.e., expected behaviors will not be visible instantly further downstream or upstream). The indices which are used to refer to each visible vehicle are as shown in Figure 71. This area of impact can be expanded by defining new indices and variables using the User Defined Attributes feature available in the recent releases of VISSIM and referencing these values in EDM.
240

Figure 71. Matrix. Area of effect with indices as reference

6) Impact

Distances

The primary trigger for initiating pullover behavior is the proximity of vehicles to the ERV. At

every timestep, the vehicles check behind them to spot the presence of an ERV. If an ERV is

detected, and the distance conditions are met (within 150 ft), the vehicle is directed to perform

a lane change to the right (Figure 72). Similarly, the action of rejoining the lane will only be

executed once a minimum headway distance of (330 ft) is satisfied.

Figure 72. Illustration. Distance limits between vehicles
7) Speeds ERV models have a custom desired speed distribution, and generally drive no more than 10 mph over the speed limit. Their desired speed is user dependent and can also violate speed limits if required.
8) Shoulder Lane behavior:
241

In the case a shoulder lane does exist in the model, then the vehicles which are pulling over from the right-most lane move over and come to a complete stop. In the PIB model, there were no shoulder lanes. Therefore, the impact of this behavior, where the vehicles pull over and stop on the shoulder, is not evaluated in this corridor.
9) Separate EDMs
Due to the clear separation of expected behaviors on the road, a decision was made to model general vehicles and the ERVs separately. This was done via two separate vehicletype specific EDMs which would function in tandem at every timestep.
Model Design: An initial outline of designing a pull over behavior model is described below:
6) The following User Defined Variables are created for the surrounding vehicles, to overcome the limit on the number of surrounding vehicles visible to any given vehicle: a. Category/Type b. ID c. Distances d. Speeds
7) Each vehicle constantly checks their surroundings for the presence of an emergency vehicle. Distance constraints are also applied.
8) If the conditions are met, further actions are controlled by the EDM. 9) Appropriate responses are applied via EDM:
242

a. Lane changes if necessary b. Speed reduction c. Complete shoulder lane pull-over and parking, if possible. 10) Once the EV has passed, minimum distance and presence checks are completed before passing the control back to VISSIM.
Pullover Frequency While a pullover maneuver should seemingly eliminate most delays that an ERV may face on its route, the expectation that all road users will adhere to such a uniform behavior is far from realistic. Each interaction between an individual vehicle and the ERV should be considered separate events with responses that are unique in their execution. The choices and responses of each vehicle are a result of several factors which vary from individual to individual, and from situation to situation.
Reluctance to yield or a lack of alertness can cause delayed responses. Drivers may also panic and cause abrupt lane changes, which may be unsafe. The differences in perception-reaction times coupled with the aforementioned human factors complicate the process of pulling over for an ERV. Environmental and surrounding factors such as traffic density, availability of space, position of vehicle in queue etc. also contribute to the complexity of the interaction. Empirical data to quantify these factors are extremely resource intensive to collect. However, ignoring the variation in pullover behaviors would result in designing for the most ideal conditions which would provide a far superior performance than what is achievable in the real-world. This study proposes the
243

use of stochastic variability as a middle ground to replicate the expected variability.
In an effort to invoke stochasticity in the models that are realistic, the variation in behavior was assumed to be a function of the speed of the vehicle, which in turn loosely relates it to the current traffic density in the corridor. We assume that the vehicles are more reluctant to pull over under denser traffic conditions and more likely to delay their pullovers under such conditions. With that assumption, a focus was placed on the frequency of pull-over for vehicles traveling under 5 kmph. For vehicles traveling faster than 5 kmph, it was assumed they would move as soon as possible. It was also assumed that everyone does eventually make a lane change for the ERV, and there is zero noncompliance. An exponential distribution curve with a mean at 2.5 seconds (Figure 73) was used to model the variation of pull-over delays in the vehicles. The curve was used to generate and assign delay values for each vehicle which meets the pull-over condition and is traveling under 5 kmph. A maximum value of 15 seconds was also used as a cutoff based on an assumption that the maximum amount of time a person would delay a pull-over is not likely to exceed 15 seconds. Although a mean of 2.5 seconds was initially selected for the modeling process, it is important to note that the execution of a lane change when suggested by an EDM is instantaneous and ignores any minimum perception-reaction time in the model. To account for the minimum time needed to perceive the presence of an ERV behind them and take an appropriate reaction, a minimum cutoff of 2.5 second was also applied to the same curve. So, vehicles take at least 2.5 seconds and at most 15 seconds to initiate a pullover after the conditions are met.
244

Figure 73. Graph. The exponential curve with the cutoff limits highlighted.
VISSIM Model Parameters The custom driver models were tested on a VISSIM model of the Peachtree Industrial Boulevard. A total of three simulation models were used as the test bed. These consisted of a no-preempt scenario and two different preemption exit strategies (In Step Exit and Normal Exit)
Each model was subjected to a total of 160 runs. These comprised of 32 runs for different seed values ranging from 1 through 5. The primary difference between each run in a particular seed was the ERV entry time. The earliest introduction time for the ERV was set to 3522.5 seconds. For every subsequent run, the entry time was incremented by 5 seconds, until 32 such entry times were simulated. This process was repeated for each seed, and for each variation.
245

Results No Pullover versus Pullover models The primary focus of the study was to reevaluate the impacts of preemption strategies by introducting realistic non-ERV driver behavior in the form of pull-over behavior to improve model accuracy. The effects of the pull-over behavior are examined under both conditions, with and without preemption for an emergency vehicle in the corridor.
The results have been compiled under the metric of travel times as shown in Figure 74. There are marked improvements across the board with the implementation of pullover behavior. A boxplot of ERV travel times constructed using the aggregated data of all the 160 runs across the various scenarios, shows a reduction of approximately 80 seconds between the two no-preempt cases, and about 40 seconds in the preemption scenario. Clearly the impact of the pullover is visibly more significant when there is no preemption. The same set of boxplots also enable us to understand the adjusted effect of preemption. The decrease in travel times of ERVs due to the introduction of preemption was on average 140 seconds when there is no pull-over. With the inclusion of the pullover capabilities, the impact of preemption diminishes to around 90 seconds of delay reduction.
246

Figure 74. Graphs. A comparison of three baseline cases without (left) and with (right) pullover behavior using ERV travel times
Since the pullover of vehicles clears the queue in front of an ERV, pull-over also contributes to the delay reduction for ERVs. Since the pull-over behavior reduces the length of queue that needs to be addressed with preemption, the net impact of the preemption is observed to be lower in the case with pull-over enabled.
Figure 75 shows the boxplots for the travel times of non-ERV vehicles in the system binned over 160 seconds intervals (1 cycle length) starting from the entry time of the ERV in the corridor. It is clearly seen that with the pull-over behavior there is a significantly higher disruption (larger delays) for the non-ERVs on the mainline in the
247

no-preemption case. However, with preemption the disruption on the non-ERVs is much lower. However in both cases, with and without pull-over, the impact of preemption on the non-ERVs tend to disappear by around the 7th signal cycle from the entry of the ERV.
Figure 75. Graphs. Non-ERV travel times observed since the entry of the ERV in the system without pull-over (Left) and with pull-over (Right)
Non-uniformity in Pullover To investigate the impact of the assumption of delay exhibited by drivers in pulling over in response to an ERV, a scenario was modeled with an idealized pull-over behavior where vehicles responded immediately to the ERV by pulling over. Figure 76 shows a comparison of the travel times of the ERV for the case without pull-over and with idealized pull-over.
248

Figure 76. Graphs. The ideal/non-realistic pullover behavior compared to the non-pullover models
The ERV travel times without preemption in the pull-over enabled scenario show a reduction of almost 170 seconds, which is better than any of the preemption cases in the pull-over disable scenario. Essentially the idealized pull-over creates a path of noresistance for the ERVs and would theoretically eliminate almost all delays experienced by an ERV, except for the time lost by the ERV while coming to a stop before crossing a red signal. Further research needs to be done where field data for trajectories of ERVs are used to fine tune the delay distribution used to replicate real world variations in pull-over delay.
249

Summary In this task, the primary focus was on the development of models that would allow for more realistic simulation of preemption scenarios than those currently accommodated by default driving behavior models in microscopic simulation platforms such as VISSIM . Although preemption strategies and their effects on Emergency vehicles have been explored by many, few studies have accounted for the complex interactions between the general traffic and the ERVs. This study developed realistic models for driving behaviors for interactions between ERVs and non-ERVs which replicate the pull-over lane clearing behaviors that are observed in the real world in response to ERVs. External Driver Models were created on VISSIM, and their effects were studied in a real-world corridor.
The ERV and non-ERV driver models were based on findings from a focus group consisting of emergency response personnel. In the absence of direct field data to calibrate the pull-over driver behavior model two variations of the model were studied, one with idealized instantaneous pull-over, and the another non-idealized conditions where the pull-over is affected by delays attributable to lack of alertness of non-ERV drivers, indecision in making the required maneuvers to yield to the ERVs, and the nonERV drivers' reluctance to give up their position in the lane over concerns of being overtaken by other non-ERV drivers leading to delays.
The idealized pull-over model essentially created a path of no resistance for the emergency vehicle by cooperatively moving out of the way of the ERV in such a way that the ERV is not required to slow down for the non-ERVs. In this scenario the only
250

delay experienced by the ERV is for its mandate to come to a full stop and checking for cross street traffic before crossing an intersection that has a red signal for the approach used by the ERV. Additional gains in travel time improvement for the ERV tended to be minimal, in the order of 10%, with preemption. To accommodate for realistic response of non-ERV drivers to an ERV, an exponential model was used to represent the distribution of the delays the drivers are expected to exhibit before pulling over. Minimum and maximum cutoff of 2.5 second and 15 seconds were used for the delays. This model showed a more realistic gain in travel times for the ERVs with preemption, in the order of 16%. On the other hand, the non-ERVs on the path of the ERV experienced a disruption due to the pull-over behavior when there was no preemption. However, with preemption, the non-ERVs on the path of the ERV tended to benefit from the preemption, with reduction in travel times for the next few minutes, due to the queue flush initiated, similar to the nopull-over scenario.
251

REFERENCES for Appendix C
Buchenscheit, A., Schaub, F., Kargl, F., & Weber, M. (2009, 28-30 Oct. 2009). A VANET-based emergency vehicle warning system. 2009 IEEE Vehicular Networking Conference (VNC),
Corts, C. E., & Stefoni, B. (2023). Trajectory Simulation of Emergency Vehicles and Interactions with Surrounding Traffic. Journal of advanced transportation, 2023. https://doi.org/10.1155/2023/5995950
How a Connected Emergency Vehicle Preemption System Works. (2019). applied information. Retrieved Nov 05 from https://appinfoinc.com/how-preemptionsystem-works/
Lidestam, B., Thorslund, B., Selander, H., Nsman, D., & Dahlman, J. (2020). In-Car Warnings of Emergency Vehicles Approaching: Effects on Car Drivers' Propensity to Give Way. Frontiers in Sustainable Cities, 2. https://doi.org/10.3389/frsc.2020.00019
Obrusnk, V., Herman, I., & Hurk, Z. (2020). Queue discharge-based emergency vehicle traffic signal preemption. IFAC-PapersOnLine, 53, 14997-15002. https://doi.org/10.1016/j.ifacol.2020.12.1998
Savolainen, P. T., Datta, T. K., Ghosh, I., & Gates, T. J. (2010). Effects of dynamically activated emergency vehicle warning sign on driver behavior at Urban intersections. Transportation Research Record(2149). https://doi.org/10.3141/2149-09
Shaaban, K., Khan, M. A., Hamila, R., & Ghanim, M. (2019b). A Strategy for Emergency Vehicle Preemption and Route Selection. Arabian Journal for Science and Engineering, 44(10). https://doi.org/10.1007/s13369-019-03913-8
Weinert, F., & Dring, M. (2015, 2015//). Development and Assessment of Cooperative V2X Applications for Emergency Vehicles in an Urban Environment Enabled by Behavioral Models. Modeling Mobility with Open Data, Cham.
Weinert, F., Dring, M., & Bogenberger, K. (2019, 27-30 Oct. 2019). Influence of Emergency Vehicle Preemption on Travelling Time and Traffic Safety in Urban Environments enabled by Innovative Behavioral Models and V2X Communication Simulation and Case Study. 2019 IEEE Intelligent Transportation Systems Conference (ITSC),
252

Zhang, L., Gou, J., Liu, K., McHale, G., Ghaman, R., & Li, L. (2009, 20-23 Sept. 2009). Simulation Modeling and Application with Emergency Vehicle Presence in CORSIM. 2009 IEEE 70th Vehicular Technology Conference Fall,
253

APPENDIX D. EVALUATION OF OPTIMAL TRANSIT SIGNAL PRIORITIZATION STRATEGIES
Introduction Background In coordinated arterial systems when buses stop to serve passengers, they typically fall out of the progression bandwidth and will likely not traverse the intersection during the current green split. Transit Signal Priority (TSP) aims to provide transit vehicles a free flow path through the intersection, or at least reduce wait time. Under the right conditions, TSP has shown the potential to reduce transit vehicle delays, improve reliability and schedule adherence, and mitigate bus bunching (Al-Sahili & Taylor, 1996; Anderson & Daganzo, 2019; Balke et al., 2000; Dion et al., 2004; Ekeila et al., 2009; Hu et al., 2014; Koonce et al., 2002; Lee & Wang, 2022; Muthuswamy et al., 2007; Rakha & Zhang, 2004; Sheffield et al., 2021).
TSP system design involves balancing tradeoffs between providing green time to the priority movement, maintaining arterial coordination, and minimizing the impact to conflicting vehicle level of service. TSP is generally implemented at a corridor level, where a transit vehicle will request priority at a given downstream intersection. The priority call is received by the downstream intersection signal controller which, depending on the projected arrival point in cycle, determines if there is a need to implement a TSP signal timing change and what TSP strategy to implement. Li et al.
254

(2008), and NTCIP Signal Control Priority Working Group. et al. (2014) provide a good reference on TSP system architectures, equipment, and concepts of operation.
TSP performance is affected by a wide range of parameters and conditions, including congestion levels, bus headways, bus stop location, detector location, green extension limit, bus arrival time within the cycle, bus stop dwell time, and TSP strategy selected. There are wide variations in the reported benefits and disbenefits of TSP. A review of the literature also shows that uncertainty in the arrival time prediction, that is, the Estimated Time of Arrival (ETA), from the point of priority request to the stop line remains a challenge in TSP implementation and effectiveness. Arrival times can vary widely depending on traffic conditions (congestion) and bus stop dwell time. Several studies report reduced TSP benefits/effectiveness as congestion increases, which may be partly attributed to increased uncertainty in travel time prediction (Ngan, 2004; Rakha & Zhang, 2004). TSP systems have been reported to be more effective for far-side bus stops (bus stop is immediately downstream an intersection) compared to nearside stops (bus stop is immediately upstream an intersection). Several simulation studies, surveys, and TSP implementation pilots reveal the need to research better strategies for TSP deployments at nearside bus stops (National Academies of Sciences Engineering and Medicine, 2020; Ngan, 2004; Rakha & Zhang, 2004). The uncertainty in ETA resulting from dwell time at near side bus stops remains a barrier to TSP effectiveness.
According to a survey by the National Academies of Sciences Engineering and Medicine (2020), current TSP field implementations have a mix of fixed location and Automated
255

Vehicle Location (AVL) detection systems. Several recent research studies, as well as a few pilot tests, have attempted to demonstrate the potential of CV technologies to improve TSP operations (Beak et al., 2018; Cvijovic et al., 2022; Hu et al., 2015; Hu et al., 2014; Hu et al., 2016; Lee et al., 2017; Mohammadi et al., 2020; Teng et al., 2019; Wang et al., 2020; Wu & Guler, 2019; Yang et al., 2019; Zeng et al., 2015). In these studies, CV data is used in ETA predictions as well as to enable improved evaluations of TSP performance. Despite the significant progress made by these studies, there is still significant effort required to reach field ready implementable TSP solutions using CV technology.
Another insight from literature is the need to further research signal timing approaches that incorporate TSP in the optimization objectives. In coordinated arterial systems, signal timing parameters including cycle lengths, splits, and offsets are normally selected to provide optimal service to the general traffic without consideration of transit vehicle's operation. This may constrain the allowable ranges for TSP parameters and limit the potential system benefits.
Problem Definition and Objectives This study uses a VISSIM simulation environment to evaluate the performance of TSP strategies and establish the critical factors and conditions that affect TSP performance. Critical items assessed are bus stop location (near side vs far side), stop dwell time, ETA, general traffic demand, and signal timing parameters. Strategies for improved consideration of transit vehicles in signal timings are proposed. Finally, given the
256

uncertainty in ETA resulting from congestion and dwell time variability, the study explores optimal TSP triggering strategies both with check-in check out (CI-CO) detectors and AVL with CV data.
Objectives 1. Evaluate TSP performance under different TSP strategies, traffic demands, transit stop locations, and signal timing parameters. 2. Assess the impact of ETA selection and settings on TSP performance for CI-CO TSP implementations. Propose an AVL algorithm to improve effectiveness under varying traffic conditions. 3. Assess the impact of dwell time magnitudes and variability, both for far side and nearside bus stops
Literature review TSP systems Transit Signal Priority (TSP) aims to provide transit vehicles with a free flow path through the intersection, or at least reduce the waiting time on the red phase. Under the right conditions, TSP has shown the potential to reduce transit vehicle delays, improve reliability and schedule adherence, and mitigate bus bunching (Al-Sahili & Taylor, 1996; Anderson & Daganzo, 2019; Balke et al., 2000; Dion et al., 2004; Hu et al., 2014; Lee & Wang, 2022; Muthuswamy et al., 2007; Ngan, 2004; Rakha & Zhang, 2004; Sheffield et al., 2021). In terms of activity, TSP systems fall under the broad categories of passive and
257

active TSP. Passive TSP involves selecting signal plans that consider and /or favor the flow of transit vehicles without active detection or tracking of transit vehicles in the system. Passive TSP is known to work best for high frequency bus routes and with good knowledge of bus speed, dwell times, schedule compliance, and other transit operational characteristics (Smith et al. (2005). Active TSP detects or tracks transit vehicles along the route and provides preferential treatment at signalized intersections by extending the normal phase green, providing early green (EG), rotating the normal phase sequence, or inserting a special bus phase. For this report unless otherwise stated, the term TSP refers to active TSP.
Active TSP can be unconditional, where all buses approaching the intersection can request and be granted priority; or conditional where priority is only provided to buses meeting preset conditions which may include schedule and headway deviations, occupancy/load, conflicts with other buses and, previous granted requests (Ding et al., 2015; Hu et al., 2015; Hu et al., 2014; Lee & Wang, 2022; Liao & Davis, 2007; National Academies of Sciences Engineering and Medicine, 2020; Zeng et al., 2015). There are numerous variations of centralized and distributed architectures. NTCIP 1211 provides a good reference for these architectures (NTCIP Signal Control Priority Working Group. et al., 2014). In distributed systems, requests are processed at the intersection level, leveraging communications between the bus and the controller. In centralized systems, TSP requests are processed at the transit or traffic management center (TMC), leveraging communications between the bus, TMC, and the intersection controller. Both centralized and distributed systems have been proposed and piloted that utilize Automated Vehicle
258

Location (AVL) to capture vehicle information including location, speed, heading, schedule deviations and other characteristics, better enabling the evaluation and processing of priority requests.
A survey carried out by National Academies of Sciences Engineering and Medicine (2020) sheds light on TSP systems implemented by various Transit agencies in US and Canada. Of 31 responding agencies, 13 operated some level of conditional TSP, while 18 agencies implemented unconditional TSP. Twenty-four of the 31 agencies use distributed TSP architectures, while 7 agencies implemented centralized systems. Seventeen out of 30 responding agencies had some level of AVL implemented in their TSP systems.
TSP Performance: Measures of Effectiveness and critical performance factors Earlier TSP studies focused on evaluating the performance of TSP under different conditions and the impact of various parameters (Al-Sahili & Taylor, 1996; Dion et al., 2004; Koonce et al., 2002; Muthuswamy et al., 2007; Ngan, 2004; Rakha & Zhang, 2004). These studies consisted of primarily sensitivity experiments using different simulation tools with hypothetical networks or actual networks as the base models. Parameters found to affect TSP performance included bus headways, bus stop location, detector location, green extension limit, bus arrival within the cycle, bus stop dwell time, and TSP strategy selected. Conditions included demand volumes for both main street and minor streets. These studies demonstrated that TSP could improve bus travel time or
259

reduce delay, with marginal impacts on cross street traffic at low demand. Disbenefits to cross street traffic increases with increasing cross street volume.

Although the potential for positive TSP benefits is quite well known, the magnitude of the reported benefits and disbenefits varies widely across studies. In part, this derives from the selection of different measures of effectiveness (MOEs). TSP evaluation normally considers impacts on both transit and other vehicles in the network. Table 8 presents a summary of some of the MOEs found in the literature. For buses, most studies used travel time or delay, averaged by a given number of buses. In addition to delay, (Ngan, 2004) used green extension success rate which is the proportion of green extensions for which the intended buses successfully passed through the intersection. For general traffic, most studies use delay but there are measurement or definitional differences across studies. For example, total or average delay over an entire period or run vs a focus on localized effects by evaluating the delay over several cycles immediately following TSP (Hu et al., 2014; Ngan, 2004). To evaluate the tradeoffs between transit and other vehicles, an increasing number of studies use person delay which considers the occupancy of both transit and passenger cars (Hu et al., 2015; Lee & Wang, 2022; Mishra et al., 2020; Zeng et al., 2015).

Table 8. Summary of MOEs in literature

Study and Methods

Transit MOEs and Results

General traffic MOEs and Results

Comments on Results

260

Rakha and Zhang Average bus

Average car delay up to 28.7% bus

(2004)

delay

Average car stops.

Simulation with INTEGRATION

Average bus stops

Average car fuel usage

software fixed location

Average bus fuel

Systemwide delay,

detector

usage

stops, and fuel use

delay decrease marginal impacts to
general traffic in most of the studied scenarios

Ngan (2004)
Simulation with VISSIMVISSIM fixed location detector

Average bus

Average cross

delay & travel

street cycle delay

time

Number of

GE success rate

recovery cycles

required

Sheffield et al. (2021)
Analyzes 3-month data from a field operational TSP system

On time performance
Mean schedule deviation.
Travel time TSP requests vs

Change in split failure
Change in green time

TSP granted

Hu et al. (2014)

Person delay in

Compares TSP with fixed check- in and a

3 cycles after TSP

modified TSP algorithm Average bus

using CV data

delay

Person delay in 3 cycles

Maximum benefits to buses at v/c = 0.8
GE less effective with increased v/c
Higher bus delays for nearside bus stops
Up to 2.5% improvement on time performance
Marginal impacts on general traffic even with unconditional TSP
CV-TSP reduced bus delay up to 88%
Delay per person in system was reduced by up to 14%

261

Ekeila et al. (2009) Average Bus

Using VISSIM, compares CTSP and a

travel time delay across

modified TSP algorithm

that uses AVL

Average Vehicle travel time and delay across simulation runs

Up to 33 % reduction in bus travel time
No significant impacts on other vehicles

Muthuswamy et al. (2007)

Average bus TT Average vehicle travel time

Up to 25% decrease in bus travel time

WATsim simultion

Wang et al. (2020) TSP requested NA

Field evaluation compared TSP

vs TSP served. bus reliability,

performance before and bus travel time

after signal retiming

bus running time

Up to 92% bus reliability

Zhou et al. (2006) Bus average

VISSIM simulation on

delay

a hypothetical

intersection and a

nearside bus stop

Total intersection delay

Bus delay reduced by up to 29%
Passenger car delay increased by up to 55%

Liao and Davis (2007)
Aimsun delay

Bus travel time Bus delay Bus stop time

Speed, average TT, 4-15% reduction in

average delay,

bus travel time and

average stops

5-20% reduction in

bus delay

Lee et al. (2017)

Bus delay Successful GE

NA

Reduced delay for the bus between 3275%

262

Field testing TSP with CV algorithm on a Virginia Tech test bed
Arrival time prediction A key challenge in TSP implementation remains the estimation of travel time on the link as the bus approaches the intersection. In both fixed location check-in detector and AVL systems, an estimate of the approaching bus travel time is needed to evaluate the need for TSP service and the appropriate TSP strategy to employ. In centralized AVL systems, a travel time estimate is also needed to know when to send the TSP request. Uncertainty in bus travel time mainly comes from the level of congestion on the link and dwell time at bus stops. There exists more extensive literature on the influence of congestion rather than dwell time. There are several studies for bus travel time prediction along the entire routes meant for schedule improvements and passenger information, but these are not sufficiently granular for TSP applications (Farid et al., 2016; Jian et al., 2013; Kumar et al., 2019; Taparia & Brady, 2021; Xu & Ying, 2017).
Several studies including Ngan (2004), Wu and Guler (2019) and Rakha and Zhang (2004) show reduced effectiveness of TSP at high levels of congestion, which is partly attributable to increased uncertainty in travel time and improperly setting ETA. There is a general adoption of historical average travel times in the setting ETA in many TSP field implementations, although limited published information or guidance exists on setting ETA. As discussed below, most of the research on ETA prediction and setting is
263

undertaken for AVL and centralized TSP systems (National Academies of Sciences Engineering and Medicine, 2020), while the larger share of TSP remains fixed location detector implementations.
Several studies that focus on developing more dynamic and adaptive TSP algorithms include ETA prediction modules (Balke et al., 2000; Ekeila et al., 2009; Kim & Rilett, 2005; Li et al., 2011). In these studies, ETA is mainly predicted by using regression models (Ekeila et al., 2009; Hu et al., 2014; Kim & Rilett, 2005; Li et al., 2011; Liu et al., 2007), Kalman filter models (Ekeila et al., 2009; Shalaby & Farhan, 2004), and more recently machine learning methods (Farid et al., 2016; Jian et al., 2013). Other studies such as Tan et al. (2008) have adopted probabilistic and Bayesian approaches. These models are built with historical and/or real time data, where the main predicting variables include the distance remaining and the dwell time for near side bus stops. Other researchers have used analytical approaches including shockwave theory to estimate queuing ahead of the bus as well as bus travel time (Bhaskar et al., 2007; Liang et al., 2018; Liu et al., 2009; Tu et al., 2012). The analytical approaches primarily use detector and signal data as their inputs. A model developed by Tu et al. (2012) uses image processing sensors to detect and process the number of vehicles ahead of the bus in real time and, together with signal status data, estimate ETA.
Koonce et al. (2002) and Zhou et al. (2006) focus not on ETA but on determining the optimal detector locations for TSP performance. Zhou et al. (2006) develops an analytical methodology for determining optimal detector placements for queue jumpers, while
264

Koonce et al. (2002) establishes a rule based method evaluated on a field TSP system. A sensitivity study by Wu and Guler (2019) found that different detector locations may be optimal depending on the level of congestion. AVL provides the capability to trigger TSP calls at any point as the bus approaches the intersection. A few studies including Liu et al. (2007) have explored the potential of AVL to find optimal trigger points for TSP which may mean different request points for each individual bus and with different strategies for GE and EG. Liu et al. (2007) demonstrates the need to request EG sufficient early such that the signal controller will be able to adjust the signal timings and clear the queue ahead of the vehicle.
Signal timing Several studies have revealed improved TSP performance given the optimization of background signal timing parameters, in particular cycle length and phasing sequences (Muthuswamy et al., 2007; Rakha & Zhang, 2004). Rakha and Zhang (2004) studied the impact of cycle length, splits, and phasing sequences on the performance of TSP and found that the vehicle optimal signal timings provide the best performance for TSP as less adjustment is needed to accommodate TSP. In one of the very few studies that report TSP field performance, Wang et al. (2020) analyzed TSP performance data for a 35 intersection CV corridor running TSP in Salt lake, Utah. The study compared TSP performance before and after signal retiming in terms of TSP requested, TSP served, bus reliability, and travel and running time. The study reports improvements on all measures after retiming. The paper, however, does not provide the before and after signal timing plans or report the changes made.
265

Starting from a max-pressure based signal optimization model developed by Varaiya (2013), Xu et al. (2022) adds transit constraints to incorporate transit vehicle operations in the base signal timings. Testing the model on a simulated Austin downtown network, the approach showed better performance in terms of bus travel time with minimal impacts to general traffic, especially at lower traffic demands. Several studies have proposed modifications to the MAXBAND signal progression model formulated by Little et al. (1981) to incorporate transit bus parameters, including speed and dwell time in the optimization of the bandwidth (Dai et al., 2015; Han et al., 2022; Jeong & Kim, 2014). Hu et al. (2015) formulates a TSP optimization model with CV data that allows communication between buses and adjacent signals to coordinate TSP and bus progression.
Incorporating CV data Several recent research studies and a few pilot tests have attempted to demonstrate the potential of CV technologies to improve TSP operations (Beak et al., 2018; Cvijovic et al., 2022; Hu et al., 2015; Hu et al., 2014; Hu et al., 2016; Lee et al., 2017; Mohammadi et al., 2020; Teng et al., 2019; Wang et al., 2020; Wu & Guler, 2019; Yang et al., 2019; Zeng et al., 2015). Despite the significant progress made by these studies, there remains significant required effort to reach CV technology field-ready implementable TSP solutions. Some of the key challenges of CV technology adoption include limited penetration, non-compatible field controllers, and communication ranges and frequencies, especially between Vehicles and the Infrastructure (Beak et al., 2018; Li et al., 2008; Smith et al., 2005; Wu & Guler, 2019).
266

A recent study by Cvijovic et al. (2022) demonstrates the state of art for incorporating CV data in TSP operations. The study formulates and tests a CV-TSP algorithm in a simulation environment in which Basic Safety messages (BSM) and Signal Request Messages (SRM) are exchanged between Onboard Units (OBUs) and Roadside Units (RSU) through Direct Short-Range Communication (DSRC). In a VISSIM simulation, these processes are implemented using the Component Object Model (COM) as a virtual server. The modified TSP algorithm showed significantly improved performance compared to the conventional TSP systems. The study estimates ETA by assuming perfect knowledge of the queuing conditions ahead of the bus. An earlier study by Hu et al. (2014) implements almost the same processes but uses a shockwave-based approach developed in (Liu et al., 2009) to estimate ETA.
CV technology provides the means to implement conditional TSP in which TSP is only granted to qualifying buses depending on the lateness, occupancy, reservice time/lockout time, and conflicting priority requests. Performance optimization models can also take CV data in which vehicle locations and occupancy are used as in person-based delay optimization models (Lee & Wang, 2022; Zeng et al., 2015).
Dwell time variability and nearside bus stops An earlier stated survey of transit agencies revealed that more than half of the current TSP deployments in the US have nearside stops (Board et al., 2020). In TSP implementation, nearside bus stops pose the challenge that dwell time, which may be highly variable, needs to be included in the ETA prediction. Dwell time can vary widely
267

across stops and at the same stop (Board et al., 2013; Isukapati et al., 2017). According to the Transit Capacity and Quality of Service Manual, TCQSM (Board et al., 2013), dwell time is associated with boarding and alighting passenger volumes, fare payment method, vehicle type and size, in-vehicle circulation, and stop spacing.
A few studies including (Kim & Rilett, 2005; Ngan, 2004; Rakha & Zhang, 2004) assess the performance of TSP at nearside stops compared to far side stops with a common result that TSP performance is better for farside bus stops compared to nearside bus stops. Only a few of these studies consider the impact of the dwell time variability, while others use unrealistic field measured dwell time distributions (Dion et al. (2014). Several simulation studies including Cvijovic et al. (2022) model dwell time using a normal distribution, although as seen in the field data presented by Isukapati et al. (2017), dwell time distributions may not be closely approximated by the normal distribution.
Methodology Overview This study uses the VISSIM simulation environment to evaluate the performance of TSP under different demand, transit vehicle and transit route parameters, and signal timings parameters. VISSIM is chosen for its ability to model: (1) TSP through its Ring Barrier Controller (RBC) module, (2) transit vehicle operation using the public transport line (PT line) submodule, and (3) connected vehicle operations using the Component Object Module as a virtual server. Additionally, the EVP component of this study discussed previously showed very comparable performance between Qfree's MaxTime
268

and VISSIM's RBC in simulating signal control and pre-emption operations. As part of this TSP effort, comparisons of MaxTime's priority settings were also seen to be similar to that of the VISSIM RBC. Therefore, in the TSP study the VISSIM RBC is taken to provide a reasonable replicate of field signal controller TSP operations. Future efforts may test critical findings in a Software-in-the-Loop environment.
The study performs experiments on variations of a hypothetical network, with the simplest being a single intersection with a far side bus stop and the more complex being a pair of two coordinated intersections with a nearside bus stop at the TSP equipped intersection. The single intersection is assumed to run in coordination, i.e., as part of a system. The initial single intersection experiments seek to test TSP outside the influence of platooned arrivals, allowing for a study of TSP responses not confounded by the arrival pattern. The two-intersection experiments added the confounding influence of platooned arrivals. Two TSP strategies are implemented: (1) both GE (green extension) and EG (early green) available and (2) only GE.
RBC- TSP Algorithm RBC's TSP algorithm functionality is described in the RBC manual (PTV, 2020). Figure 78 is an illustration of the RBC's TSP algorithm in coordination. In the figure Check-in is the time in the cycle when the TSP call is placed and Y+AR is the yellow plus all red for the given phase.
The algorithm operates by granting green extension (GE) or red truncation/early green (EG) to the priority movement or maintaining the current phase timing. The GE, EG, or
269

no priority timing decision is based on the receipt of a call from the intersection's upstream check-in detector and the estimated time within the cycle that the bus would arrive at the stop line. Key inputs include Estimated Time of Arrival (ETA) and maximum GE. ETA is a user input of the estimated/predicted time for the bus to travel from the check-in detector to the stop line. The maximum allowable GE is limited, such that in the cycle after the TSP service there is sufficient time to serve the specified minimum green for each movement. The user can specify any desired maximum GE value below the algorithm's maximum allowable GE.
Upon check-in, the TSP algorithm determines the projected point in the cycle when the bus will arrive at the intersection. If the bus is projected to arrive at the stop line during the phase green time, no action is taken, i.e., the end of green is not altered. If the bus is projected to arrive after the phase green time end, but within the limits of specified maximum GE, GE is granted until the bus checks-out or maximum GE is reached. If the bus is projected to reach the stop line before the start of the phase green time (i.e., arrive on red but beyond point where GE could be applied), EG is requested. Ideally the algorithm will start truncation at a point such that the bus arrives at the stop line without needing to stop. The EG allowed is limited to the remainder of the preceding phase green after serving the maximum of the phase minimum green or a preset minimum called priority minimum green. This minimum green can be the pedestrian time for through movements, if applicable. The algorithm also allows the user to set phases that may be skipped to reach the priority phase earlier. The user can control the algorithm and balance the tradeoff between bus travel time and general traffic delay by (1) selecting appropriate
270

priority minimum greens, (2) selecting appropriate maximum GE, (3) specifying the phases that can be skipped, if any, and, (4) setting the reservice parameter which specifies the minimum time to grant the next TSP request.
Figure 77. Matrix. TSP Algorithm
TSP Parameter settings As mentioned, the TSP algorithm requires user inputs for several parameters. Three of the most critical are ETA and maximum GE and EG. In all experiments in this study,
271

maximum GE was limited to 20% of the cycle length based on common practice observed in the literature. The base experiment TSP allows both GE and EG options. Later experiments test the GE only strategy as this approach was undertaken by a subset of agencies in the literature. Minimum greens are set to allow no truncation of the coordinated movement. For side street movements, maximum truncations of 5 and 10 seconds are tested. On the main street 5 second truncation is set for the two left turn movements in all experiments.
Check-in detector location (TSP trigger location) is limited to the distance that can be traveled by the bus at free flow speed. In the far side bus stop experiments, the check-in detector is located after the upstream intersection far side bus stop which becomes a limiting constraint in cases where maximum allowable GE is greater than free flow travel time from the bus stop. This will limit the number of buses that would potentially request GE.
The VISSIM RBC, similar to many field controllers in use today, allows only a single value input of ETA. This value is consistent across all time-of-day plans and free operation. Thus, for all buses in each simulation run, a single value of ETA is set. While this may not be optimal, numerous field deployed TSP systems with fixed location detectors employ the same approach. Objective two of this study focuses on the importance and selection of this parameter.
272

Test Network Experiments are performed on hypothetical networks consisting of one or two intersections, utilizing actuated coordinated signal timing. As shown in Figure 78, the base network model consists of a bus route on the main street (E/W) with the bus running in the coordinated movement. There is a fixed location check-in detector 1000ft upstream of the stop line, a checkout detector immediately after the stop line and a far side bus stop immediately downstream of the intersection. The network has two through lanes in each direction for the major street and one lane in each direction for the minor street (N/S). All left turns on the major and minor streets have exclusive turn lanes and only protected signal phases. For simplicity, right turning movements are omitted for both streets.
Figure 78. Illustration. Base network model in VISSIM, Far Side Bus Stop
For the dwell time experiments of objective three, an adjacent upstream intersection is added to the model as shown in Figure 79 to enable simulation of platooned arrivals. With arrival profiles controlled from the upstream, the impacts of bus stops including dwell time variability may be more easily be isolated and evaluated.
273

Figure 79. Illustration. Two intersection network model
Signal timings and Traffic Demand The base model considers actuated-coordinated signal timing with a cycle length of 100 seconds and splits as shown in Table 9. Each phase has a yellow time of 4 seconds and an all-red time of 1 second. Floating force offs are set to allow unused green to go to the main street coordinated through movements. For the experiments, volumes are calculated to achieve the desired overall intersection volume to capacity (v/c) ratio. Both the major and minor streets have balanced volumes in all corresponding opposing directions, for instance Eastbound through (EBT) volume is equal to westbound through (WBT) and Eastbound left turn (EBL) volume is equal to westbound left turn (WBL) volume.

Table 9. Signal Timing Parameters

Signal Group (SG) 1

splits

15

Yellow

4

All-Red

1

Min Green

5

2

3

45 20

4

4

1

1

20

5

4

5

6

7

8

20 15 45 20 20

4

4

4

4

4

1

1

1

1

1

10

5

20

5 10

274

Preliminary experiments were run in VISSIM to extract saturation headways such that matching values are used in calculating volumes for the different degrees of saturation. Experiments were conducted for v/c ratios of 0.85, 0.95, and 1.0. The volumes to achieve the v/c ratios are given in Table 10. In the movement column, the first two characters of each entry represent network direction while the last character represents movement direction. The volumes are in veh/hr.

Table 10. Volumes for different v/c ratios

Movement
EBL EBT WBL WBT NBL NBT SBL SBT

Volume for v/c= 1.0 (vph) 180 1515 180 1515 270 284 270 284

Volume for v/c = 0.95
(vph) 171 1440 171 1440 257 270 257 270

Volume for v/c = 0.85
(vph) 153 1288 153 1288 230 241 230 241

275

To capture sufficient bus samples, each simulation scenario is run for 30 hours with a bus entering approximately 20 minutes into in each hour (additional detail in the next section). To ensure independent data for each bus arrival volumes were entered in VISSIM in intervals of 15 minutes with the last 15 minutes of each hour assigned a minimum flow of 200 veh/hr and 50 veh/h for the main street through movement and all other movements, respectively. Reducing the volumes in the last 15 minutes allows any accumulated queues or other impacts from the prior potential bus priority to clear the network. A minimum of 15 minutes is allowed for warm up at the start of each hour before collecting any data.
Bus inputs Transit buses enter the system from the beginning of the main street on the west side and drive eastward through the intersection(s). VISSIM provides the capability to model transit vehicle operations including routes, bus stops, dwell time, passenger loads, and passenger flows through the modules of PT lines and PT line stops. One bus enters the system per hour to allow isolation of the impacts of each TSP event. Each run of each experiment lasts for 30 hours to allow a sufficient sample of TSP events.
To capture different TSP request scenarios in the isolated intersection experiments, bus arrivals were approximately uniformly spread across the cycle length. To achieve this, each bus is introduced after a warm-up period and an additional random term between zero and the cycle length. Figure 80 shows the resulting number of buses arriving at the
276

check-in detector at a given point in cycle. The data includes 300 buses, generated from 10 replicate trials of 30 hours each, with one bus entry each hour. All experiments in this study consider unconditional TSP in which all arriving buses request and can be granted priority without fulfilling any preset conditions.
Figure 80. Graph. Arrival profile of buses vs time in cycle
Experiment design Overview
As per the objectives of this study, three sets of experiments are designed. The first set is designed to evaluate critical parameters and conditions that affect TSP performance. Under this category the first experiment analysis includes: (a) comparing the GE and EG TSP strategies, (b) testing TSP performance at different levels of traffic demand, and (c) assessing the impact of cycle length on both transit vehicles and general traffic. The second set of experiments studied: (a) travel time variability resulting from congestion on
277

the bus route intersection approach, (b) considerations in setting the ETA, and (c) leveraging the capabilities of AVL and CV data to devise improved TSP triggering mechanisms. The third set of experiments are performed on the two-intersection model and are designed to study dwell time variability on TSP performance for both far side and near side bus stops. While not directly tested, the importance of the headway between buses will also be deduced in the experiments through consideration of number of cycles impacted by a bus traversal of an intersection.
Experiment Set 1
The first experiments: (a) compare the GE and EG TSP strategies, (b) test TSP performance at different levels of traffic demand, and (c) assess the impact of cycle length on both transit vehicles and general traffic. During each simulation run, performance data including travel time and delay (bus and general traffic traffic), as well as high-resolution controller data consisting of detector actuations, signal changes, and TSP events are collected and archived for post processing. Second-by-second controller data allowed for assessing the effectiveness of TSP, including time of check-in and check-out, shares of GE and EG, proportion of successful/effective GE, length of GE, amount of truncation on each phase, and timing changes in the recovery cycle/cycles. Unsuccessful or ineffective GE refers to when GE is granted and maximum GE is reached before the bus reaches the stop line. When considering delay up to 18 cycles (30 minutes) after check-in were investigated for differences between no TSP and TSP. In post processing run results, controller data was fused with performance data to isolate and define the impacts of each TSP event.
278

(a) Experiment 1: Comparing GE and EG with GE only TSP strategy
As mentioned, some transit agencies operate both GE and EG in their TSP implementations while others opt for only GE. Experiments under this sub task are therefore designed to evaluate the benefits and disbenefits of the two TSP strategies, (1) GE and EG, and (2) GE only. The RBC settings allow the user to select either of the two strategies of GE and EG or GE only. For the base model with v/c= 0.95, 10 replicate simulation runs were performed for each of the strategies and the baseline no TSP case.
(b) Experiment 2: TSP performance under different demand levels
This experiment was designed to assess the performance of TSP under different levels of congestion. Starting with a base model with an overall intersection v/c ratio of 0.95, volumes are adjusted while maintaining the same signal timings (cycle length and splits) to achieve overall v/c ratios of 0.85 and 1.0, as shown in Table 10. At each v/c ratio new optimal ETA values are set in the controller to account for varying travel time. It was hypothesized that (1) at lower v/c ratios TSP may result in a higher rate of successful GEs and (2) at higher v/c ratios, GEs and EGs may not only have larger disbenefits to side street traffic but may also be less effective in serving transit vehicles due to the increased uncertainty in travel time and higher likely of an unsuccessful GE.
(c) Experiment 3: Impact of cycle length
Cycle length is normally selected to minimize general traffic delays and to maintain coordination in arterial systems. This is normally done without considering transit vehicle
279

operations. This experiment is designed to assess the impact of cycle length selection and the tradeoff between general traffic delay and TSP efficiency. It was hypothesized that higher cycle lengths could improve TSP efficiency by providing more flexibility in terms of the available maximum allowable GE and EG /truncation limits and reduce the potential negative TSP impacts (i.e., increased delays) to the side street. However, higher cycle lengths may increase overall intersection delay when under normal operation.
With the v/c ratio of 0.95 volume set in Table 10, an optimal cycle length of 110 seconds is found to minimize general traffic delay. The corresponding signal group splits are shown in Table 11. For all signal groups (SG), a total clearance time (Y+AR) of 5 seconds is provided. For the 110 second cycle, a maximum allowable GE set to 20% of the cycle length is equal to 22 seconds. For EG, minimum priority green, which is the green that must be served before truncation, is set as shown in Table 11. A truncation of up to 10 seconds is allowed on side street phases 4 & 8 and 3 & 7, while a truncation of 5 seconds is allowed on the main street left turns, i.e., 1 & 5. No truncation is allowed on the main line through movement. For example, signal groups 4 & 8 have a split of 22 seconds with a clearance of 5 seconds, thus 17 seconds of displayed green. Allowing a maximum truncation of 10 seconds leaves the minimum priority green as 7 seconds.
Cycle lengths of 130 and 150 seconds, as shown in Table 11, are selected to test TSP performance when the cycle is above the general traffic delay optimal length. To have similar GE and EG amounts or ranges, the maximum allowable GE and EG limits are same as for the 110 second cycle length. For EG, maintaining the same allowable
280

truncation limits, the corresponding minimum priority green durations for the cycle lengths of 130 and 150 seconds are shown in the table. For each cycle length, 10 replicate simulation runs are made for the TSP and no TSP cases, resulting in a total of 60 simulation runs.

Table 11. Splits and maximum GE for varying cycle lengths

cycle length SG

1 2 3 4 5 6 7 8 max GE

Splits

16 50 22 22 16 50 22 22 22

110

Min priority green 6 45 7 7 6 45 7 7

Splits

19 59 26 26 19 59 26 26 22

130

Min priority green 9 54 11 11 9 54 11 11

Splits

21 70 30 29 21 70 30 29 22

150

Min priority green 11 65 15 15 11 65 15 15

Experiment Set 2 The second set of experiments study: (a) setting the ETA and (b) leveraging the capabilities of AVL and CV data to devise improved TSP triggering mechanisms.
(a) Estimated Time of Arrival (ETA)
One of the key challenges and determinants of TSP effectiveness is the estimation of ETA, the travel time from the point of priority request to the stop line. Uncertainty in ETA may come from congestion on the link or from dwell time at near-side bus stops. This task focuses on ETA uncertainty resulting from congestion and is therefore carried
281

out on the model with a far side bus stop. The first part of the experiment assesses the variability of ETA at three congestion levels, represented by v/c ratios of 0.85, 0.95 and 1.0. It is hypothesized that ETA variability increases with the degree of saturation. In the current versions of the RBC and MaxTime controllers ETA is a fixed value and therefore a single value must be selected to cover both peak and off-peak hours.
Limited information is available in the literature regarding ETA value selection, except the mention that ETA is estimated as an average of historical travel times. In this study, simulated "historical" bus travel times are analyzed for variability. In post processing the historical travel times, only buses that check-in on green and proceed through the intersection on the same green or its extension are included in the ETA setting analysis. This eliminates travel times that have segments with the bus waiting at a red indication. The considered trajectories include two cases, (1) the bus passes the check-in detector and proceeds to pass the stop line without stopping and (2) the bus checks in on green, stops in queue between detector and stop line but passes through the intersection on the same green time. The simulated "historical" data is generated from 10 replicate 30-hour simulation runs for each congestion level.
After establishing the variability of ETA with varying degrees of saturation, the second part of the experiment performs a sensitivity analysis with different ETA values guided by the observed variability in the first part of the experiment. Experiments are performed for only the v/c ratio of 1.0 and 10 replicate runs are performed for each ETA.
(b) TSP with AVL system 282

AVL systems provide the capability to place priority calls at any point upstream of the stop line as opposed to fixed location detector systems. For instance, during heavy congestion, for the same ETA setting in the controller, priority calls could be placed closer to the stop line. Similarly, a priority call could be further from the stop line in light congestion conditions. Thus, while the ETA remains a fixed value, the location at which the bus places the call is flexible. Additionally, AVL systems provide the means to update or cancel calls if the ETA estimate changes significantly (NTCIP Signal Control Priority Working Group. et al., 2014). It is increasingly common that many transit buses are equipped with some level of AVL system capable of tracking and transmitting at least bus location, heading, and speed to the transit or traffic management center (TMC). A survey by National Academies of Sciences Engineering and Medicine (2020) found more than half of the responding transit agencies have some level of AVL in their TSP implementations. The capability of AVL systems is partly determined by the means and frequency of data exchange between the bus onboard GPS system, the signal controller, and the TMC.
In this task, a modified AVL based TSP algorithm is formulated and tested in VISSIM simulation environment. It is assumed that all buses are equipped with onboard GPS units transmitting their location and speed at a second-by-second frequency. In this simplified experiment, perfect knowledge of the traffic conditions ahead of the bus is assumed. This occurs where high CV penetration exists in the general traffic or extensive external detection is in place. In future studies this assumption shall be relaxed. In the current implementation, as the bus approaches the stop line the ETA is updated at every time step
283

as a function of bus speed (v), distance remaining (D), and number of vehicles ahead of the bus (n). ETA is predicted as the maximum of the time to discharge vehicles ahead of the bus, estimated as (hs*n) where "hs" is the saturation headway, and the time to traverse the remaining distance on the link (D/v), where "v" is the prevailing bus speed. This covers the two cases of the bus stopping at back of queue (hs*n) or proceeding at the prevailing speed without stopping (D/v).
= max( , )
In addition to estimating and tracking ETA, the algorithm tracks the signal state. In field implementations, this would be in form SPaT data, available at a minimum of a 1 Hz frequency. The algorithm uses different strategies to trigger a priority call depending on whether the signal state is red or green, anticipating EG or GE respectively. For EG, the call needs to be placed with sufficient time for the signal to implement EG. When GE is anticipated, the objective is to guard against inefficient green extensions while not sacrificing successful GE opportunities.
For the VISSIM implementation, an ETA equal to max GE is set in the controller. For instance, in this experiment both max GE and ETA are set as 20 seconds. Requests for both GE and EG are only sent after the bus has passed 1000 ft, at which it is assumed that the bus has cleared any upstream bus stops. When the bus has passed the 1000ft mark and the signal state is red, a request for EG is triggered. When the bus has passed the 1000ft mark and the light is green, a priority request for GE is sent when predicted ETA falls
284

below the ETA set in the controller. This differs slightly from the TSP system architectures described by NTCIP 1211 v02 of 2014 in which a unique ETA for each bus is sent to Priority Request Server (PRS). This work around is partly needed because of the limitation that ETA cannot be changed/updated during a simulation run in either VISSIM's RBCs and Qfree's Maxtime.
The modified algorithm is implemented using VISSIM's Component Object Model that allows fetching and feeding data into the model during simulation runtime. TSP requests are placed by taking advantage of the manual detector actuation capability afforded by VISSIM. The modified algorithm is tested at a v/c of 1.0, where the normal check-in fixed detector system is challenged by the highly variable travel time. The performance of the modified algorithm is compared to the performance of the conventional check-incheck-out TSP system. Each test scenario is replicated 10 times with different random seeds. The TSP baseline case, GE and EG strategy, and GE only requires 30 simulation runs, 10 replicates for each. Adding 30 runs for the fixed detector check-in results in a total of 60 simulation runs. As previously, each run is 30 hours with a bus entering the system at each hour.
Experiment Set 3 In coordinated arterial systems when buses stop at a bus stop to serve passengers, they typically fall out of the progression bandwidth. Depending on the dwell time magnitude the buses may not be able to traverse the intersection during the current green. In a fully coordinated and under saturated system, the magnitude and variability of dwell time may largely determine the bus arrival profiles at the stop line. In this effort, experiments are
285

designed to (1) assess the impact of magnitude and variability of dwell time on TSP performance and (2) assess the potential of improving TSP performance by selecting signal timing offsets that consider bus dwell time and evaluate tradeoffs for the general traffic.

Dwell time magnitude and variability at far side bus stop
This experiment is performed on the two-intersection model of Figure 79. TSP is implemented on the East intersection, for the eastbound through movement. Each intersection models the same volumes and same signal timings, except for the TSP settings. As shown in Table 12, the volumes are adapted with minor modifications from the single intersection v/c = 0.95 experiment. For each intersection, base mode signal timings have a 100 second cycle length with splits as shown in Table 9. For this first experiment, an offset of 22 seconds is selected to provide full bandwidth for the EB priority movement.

Table 12: Two intersection model volumes

EBL EBT WBL WBT NBL

East Intersection 171 1440 171 1354 257

West Intersection 171 1354 171 1440 257

286

NBT

270

270

SBL

257

257

SBT

270

270

Dwell time data To ensure the variation in dwell time reflects potential field conditions Automated Passenger Counts (APC) data, provided by Metropolitan Atlanta Rapid Transit Authority (MARTA), was obtained. Data was available for the three busiest transit routes, 39, 05 and 83, collected over a period of one year, from August 2021 to August 2022. Exploratory analysis of the data showed dwell time variability differed across stops. For majority of the stops there was variability by passenger activity, time of day (TOD), schedule deviation, etc. Figure 81 shows the Cumulative Density Function (CDF) plots of dwell time for different stops on Route 39 Southbound in the AM peak. Each curve represents the CDF for a stop. As seen from the figure, moving from left to right, dwell time increases both in magnitude and variability.

287

Figure 81. Graph. Dwell time Distributions from field data
The dwell time magnitudes presented in Figure 81 do not include deceleration and acceleration into and out of the stops. Three dwell time distributions DT01, DT02 and DT03, representing low, medium, and high dwell time ranges were selected for the experiment. In VISSIM the distributions are input directly as CDFs. Figure 82 shows dwell time distribution DT02 as entered in VISSIM. VISSIM utilizes the provided distribution to stochastically assign each transit vehicle a dwell time at the bus stop, with the bus skipping the stop if the assigned dwell time is zero. Analysis of preliminary VISSIM run outputs showed very similar dwell time plots as that of the field data.
288

Figure 82. Graph. Dwell time distribution 02 input in VISSIM
This set of experiments is performed on a far side bus stop with the check-in detector positioned after the bus stop. Dwell time is therefore not included in the ETA prediction, although dwell time impacts the arrival time of the bus at the downstream intersection. For each dwell time distribution, 10 replicate runs are performed with TSP activated and 10 replicate runs with TSP deactivated on the East intersection.
(a) Impact of dwell time variability at nearside bus stops
As discussed in the literature, uncertainty in ETA is compounded by the variability of bus dwell time at nearside bus stops, where a TSP request is sent to the controller before the bus passes the bus stop and without knowledge of whether the bus will stop or the stop duration. The challenges introduced by this uncertainty includes: (1) providing GE but the bus dwells at the stop longer than expected and GE time is ineffective, (2) providing EG when it is not required as the bus is delayed at the stop, and (3) not requesting TSP
289

anticipating bus dwelling that does not happen. The first two are detrimental to both buses and side street traffic while the last affects only the bus.
Figure 83. Illustration. Nearside bus stop model in VISSIM
The model is constructed as shown in Figure 83, with a curb side bus stop 100ft from the stop line and a check-in detector maintained at 1000ft from the stop line. In this arrangement ETA estimates must include bus travel time at the prevailing speed, dwell time at the bus stop, and bus deceleration and acceleration at bus stop. This experiment is performed with the same volumes and signal timings as in the previous two intersection model. Preliminary runs are performed to estimate ETA for each dwell time distribution, which similar to estimating bus travel time from historical data. In this case ETA includes the dwell time at the bus stop. Distribution DT01 gave an average travel time of 42 seconds with a standard deviation of 6 seconds, distribution DT02 gave an average travel time of 46 seconds with a standard deviation of 6 seconds, and distribution DT03 gave an average travel time of 48 seconds with a standard deviation of 8 seconds. For each distribution, experiments were run several estimates of
290

ETA to allow for studying the impact of the ETA value. For comparison purposes, the experiments are repeated with TSP deactivated.
Results TSP MOEs and Critical performance parameters and conditions
Measures of Effectiveness The selected MOEs included (1) proportions of granted GE and EG, and proportion of effective GE (i.e., the bus successfully traversed the intersection), (2) transit bus travel time, (3) general traffic delay, and (4) extent/propagation of the delay after the TSP event. The first MOE is mainly presented with tables while the last three are presented with box plots to capture the variability in travel time for different buses and in the resulting delays to the other vehicles.
General traffic delay Figure 84 presents the delay for the cross street southbound left turn (SBL) movement after a 10 second truncation of the movement green, where the mainline received a 10 second EG. The intersection v/c ratio is 0.95. In the figure (and subsequent similar figures), lower and upper limits of the boxes represent the 25th and 75th percentiles respectively, the red squares represent the mean values, the horizontal line through the box is the median position, and the whiskers represent 1.5 time the interquartile ranges.
It is seen that compared to the no TSP case, delay increases and is only able to recover after approximately eight cycles. This figure represents the worst case, i.e., ten second truncation, as ten seconds is the maximum allowable truncation on the movement.
291

However, not all buses required the max allowable truncation. Where buses received varying levels of EG and GE, resulting in varying levels of truncation of the SBL movement, on average four (4) cycles after each TSP event was sufficient for recovery.
Figure 84. Graph. SBLT delay extent after 10 second EG
Comparing impacts of GE and EG Figure 85 show the impact of GE and EG, on bus travel time for a v/c ratio of 0.95. Each simulation run represents 30 bus arrivals. Comparisons between TSP and no-TSP are made for the same bus. That is, if the third bus in the simulation run receives EG, then the third bus in the no-TSP will be used for comparison. It is seen that GE results in higher levels of improvement over no-TSP than EG. This is intuitively reasonable as GE results in a bus skipping the entire red duration whereas as EG reduces the amount of red by the length of the EG truncation.
292

Figure 85. Graph. Impact of GE and EG on bus travel time at v/c ratio of 0.95
Figure 86 shows a comparison of SBT cross street general traffic delay related to (a) GE and (b) EG. The delay is calculated over 4 cycles after the TSP event. The magnitude of the change in delay (between TSP and no TSP) resulting from EG is visibly higher than that from GE. This is to be expected as the impact of truncation is borne by the one movement or movements which have been reduced to provide the priority movement, i.e., return to green more quickly, while the impact of GE is distributed across all movements in the next cycle after TSP.
293

(a)

(b)

Figure 86. Graphs. Cross street SBT Delay at v/c ratio of 0.95 for (a) EG and (b) GE

Figure 87 shows the extent of delay after (a) 10 second GE and (b) 10 second EG. The impact of EG extends at least eight cycles while the impact of the GE dissipates after the four cycles. As discussed above, the 10 second truncation due to EG is borne by the SBT movement alone while the 10 second GE is spread proportionally to all movements according to the splits. Moreover, with the split for the priority movement (main street through) being higher than the cross-street movement, most of the GE impact is taken by the main street through movement in the next cycle.

294

(a)

(b)

Figure 87. Graphs. SBT delay extent with v/c ratio of 0.95 after 10 second (a) GE and (b) EG

It is seen in Figure 85 through Figure 87 that the GE has higher bus travel time benefits than EG, while EG results in more significant side street delays. While not shown, similar trends could be seen for the v/c ratios of 1.0 and 0.85.

TSP performance under different demand levels Table 13 shows the TSP strategies and TSP effectiveness for v/c ratios of 0.85, 0.95 and 1.0. In the first TSP strategy both GE and EG are permitted while in the second only GE is implemented. "No Action" is the number of buses for which no TSP was requested, which means either that the bus could make it through the intersection on the normal green phase or the settings prevented the request. "EG" indicates the number of buses that received Early Green while "GE" indicates the number of buses that received green extension. "GE_Eff" shows the number of buses that received green extension and successfully traversed the intersection on the provided GE. GE_NoEff represents the number of buses that were granted GE but maximum GE was reached before the buses checked out/moved through the intersection.

Table 13: TSP strategies and Effectiveness under different levels of demand

TSP Strategy v/c

EG

GE & EG 0.85

119

GE GE_Eff GE_NoEff No_Action

62

62

0

119

295

0.95

120

105

100

5

75

1

129

123

100

23

48

0.85

0

62

62

0

238

GE only

0.95

0

109

104

5

191

1

0

134

111

23

166

From the table the following can be noted:
(1) From the No_Action and GE columns, more buses need TSP as the congestion increases. This is expected as at lower congestion levels there is less queuing, higher speeds on the mainline, and additional green for the coordinated mainline movement resulting from side street movements gapping out.
(2) Green extension effectiveness decreases with congestion as indicated by number of failures (GE_NoEff). For example, 23 out of 134 and 23 out of 123 green extensions for the v/c ratio of 1.0 are ineffective for the two cases of GE only and EG & GE. As discussed later, GE failures are a result of the bus having higher than the set ETA.
Figure 88 shows the bus travel time under the combined strategies of (a) GE and EG and (b) from implementing GE only. Only the buses that receive TSP are included in the results. A higher bus travel time is seen as the congestion level increases from v/c ratio of 0.85 to v/c ratio of 1.0 in the baseline no TSP case. The variability in travel time also increases as congestion levels increase. For v/c of 0.85 and 0.95, GE reduces the travel time for this subset of buses to almost free flow travel time, while for the v/c ratio of 1.0
296

several vehicles receiving GE still face queuing and therefore the average travel time is higher than the free flow travel time. Moreover, the failed GE buses sit through the entire red period and, as seen in Table D8, the number of failures increases with travel time. As mentioned already, EG provides modest benefits compared to GE and thus averaging across the buses receiving both GE and EG (Figure 88 (a)) shows comparatively lower benefits versus GE only (Figure 88(b)).

(a)

(b)

Figure 88. Graphs. Impact on bus travel time at three levels of v/c ratio, for (a) GE and EG and (b) GE only, on Bus travel time cross v/c
Side street Delay Figure 89 presents the cross street SBL and SBT movement delay with GE and EG implemented. To allow for comparison across v/c ratios delay is measured for four cycles, starting with the check-in cycle. As expected, the higher v/c ratios of 0.95 and 1.0 show higher delay increases compared to the v/c of 0.85. With a v/c ratio of 0.85 the additional delay is practically insignificant. Additionally, when considering later cycles
297

(five or more cycles after TSP) additional delay to the side streets remains for higher v/c ratios, with the impact of TSP dissipating more slowly than for lower v/c ratios as will be seen in the next set of figures.

(a)

(b)

Figure 89. Graphs. Cross Street Delay from GE and EG for (a) SBL and (b) SBT

Figure 90 presents the cross street SBL and SBT movement delay with only GE implemented. For both movements and for all v/c levels, the delay increases slightly. This reinforces the earlier observation that EG will cause higher disbenefits to the side streets compared to GE.

298

(a)

(b)

Figure 90. Graphs. Cross Street Delay from GE only for (a) SBLT and (b) SBT

Side street Delay Extent

The three v/c ratio demand levels are further compared in terms of the extent of delay

caused by different levels of EG to the side street movements. In the results below the

TSP delay extent for each movement is presented as (1) the overall delay extent caused

by any EG up to the maximum, and (2) the delay extent caused by only the maximum set

truncation. For example,

(b)

Figure 91 (a) shows the extent of SBL movement delay, at the number of cycles past

TSP, caused by any truncation up to 10 seconds, while (b) shows the delay extent caused

by only those truncations of 10 seconds. That is, the delay in part (a) averages the impacts

of different truncation amounts ranging from 0 to 10 seconds. While the data is for only

buses that proceed on EG, there are cases where the subject movement is not truncated

(e.g., only SBT is truncated whereas SBL is not) and hence some side street vehicles are

299

included in the data that do not experience any changes in delay from non-TSP to TSP cases. For the v/c ratio of 1.0, Figure 91 (a) shows the SBL movement delay change persisting up to the 7th to 9th cycle while Figure 91 (b) shows the additional delay persisting up to at least the 13th cycle, with more significant impacts on the variability as well. As expected, the maximum truncation only ( Figure 91(b))has a higher impact throughout the duration after TSP. Thus, as truncation increases both the impact level and duration increase.
(a) Degree of saturation, v/c = 1.0

(a)

(b)

Figure 91. Graphs. SBL Delay Extent, v/c = 1.0 for (a) 0 to 10 second truncation and (b) 10 second truncation only
Figure 92 presents the results for the SBT movement. In both part (a) and (b) the delay persists through the 14 observed cycles, indicating that SBT experiences more delay increases compared to SBL. This is due to the position of the two movement phases in the phase sequence, with SBT likely to be truncated by a higher amount compared to

300

SBLT. Figure 92 (a) and (b) are also more similar than (b) Figure 91 (a) and (b) as most EG provisions reach the set maximum of 10 seconds. Thus, the main finding is that for the most affected movement, SBT, under v/c ratio conditions of 1.0, the delay increase is significant and persists through the observed 14 cycles after truncation for the set maximum truncation of 10 seconds and v/c= 1.0. While slightly muted the same trends are seen when the maximum truncation is limited to 5 seconds (figures not shown).

(a)

(b)

Figure 92. Graphs. SBT Delay Extent, v/c = 1.0, (a) 0 to 10 second truncation and (b) 10 second truncation only

(b) Degree of saturation, v/c = 0.95 Figure 93 and Figure 94 show the delay extent for SBL and SBT movements respectively at the v/c ratio of 0.95. For both movements, delay persists up to the 8th cycle, recovering earlier than in the v/c ratio of 1.0. Additionally, while significant side street delay does result from the TSP, it is lower than seen in the v/c ratio of 1.0. This is as expected as the 0.95 v/c ratio allows for some slack in the timing to recovery from the TSP. Again

301

similar, although more muted, results were seen when maximum truncation was set to 5 seconds (figures not shown).

(a)

(b)

Figure 93. Graphs. SBL Delay Extent, v/c = 0.95, (a) 0 to 10 second truncation and (b) 10 second truncation only

(a)

(b)

Figure 94. Graphs. SBT Delay Extent, v/c = 0.95, (a) 0 to 10 second truncation and (b) 10 second truncation only

(c) Degree of saturation, v/c = 0.85 Figure 95 and Figure 96 show the delay extent for SBL and SBT after TSP events for a v/c of 0.85 and maximum set truncation of 10 seconds. For SBLT delay increase is only observed in the first three cycles considering all truncation amounts up to 10 seconds,

302

with the magnitude of the delay increase well below that of the v/c ratios of 1.0 and 0.95. Only five cycles were required for recovery when considering only truncations of 10 seconds. For SBT delay increase is observed in the first five cycles considering all truncation amounts and up to six cycles for truncations that reach the set maximum of 10 seconds. Again, similar but more muted results were seen when the maximum truncation was set to 5 seconds. The trend in these results demonstrates the critical importance in having slack in the signal timing plans (i.e., v/c sufficiently below 1.0), where the desire is to limit the impact to non-TSP movements.

(a)

(b)

Figure 95. Graphs. SBL Delay Extent, v/c = 0.85, (a) 0 to 10 second truncation and (b) 10 second truncation only

303

(a)

(b)

Figure 96. Graphs. SBT Delay Extent, v/c = 0.85, (a) 0 to 10 second truncation and (b) 10 second truncation only

304

Impact of cycle Length on TSP Performance TSP Effectiveness Table 14 shows the TSP strategies and effectiveness at different cycle lengths. As expected, the number of buses not requiring TSP increases slightly as the cycle length increases. This is due to increased mainline green. Furthermore, as the cycle length increases the number of buses proceeding on GE reduces while the number of buses requesting EG is higher. This proportion of GE & EG partly determines the TSP efficiency, with higher GE preferred.

Table 14. TSP strategies at effectiveness at different cycle lengths

cycle C110 C130 C150

EG

GE

108

87

125

58

125

45

GE_Eff GE_NoEff No_Action

87

0

105

58

0

117

45

0

130

Bus Travel Time Figure 97 compare bus travel times with and without TSP for the three cycle lengths. In Figure 97 (a) only TSP affected buses are included in the analysis while in Figure 97 (b) all buses are included. The maximum TSP benefit to buses that receive TSP is provided by the optimal cycle length of 110 seconds. This is likely due to increasing percentages of EG and decreasing percentages of GE as the cycle length increases, as seen in Table 14. When considering all buses, the 110 second cycle continues to provide lower bus travel times, although the differences in bus travel time between cycles lessens. Note that this part of the experiment constrains maximum GE and EG for the two higher cycles to that
305

of the optimal cycle and thus there is potential for more travel time savings with the two higher cycles.

(a)

(b)

Figure 97. Graphs. Bus travel time vs cycle length for (a) Only TSP affected buses and (b) All buses

Cross street Delay Figure 98 (a) and (b) shows the cross street SBLT and SBT delays respectively, for only those bus traversals where TSP was provided. Delay is averaged for 4 cycles after TSP. As expected, in the no TSP case delay increases as the cycle length increases from the optimal cycle length. With TSP, the change in delay after TSP service is lowest at 130 seconds. This reinforces the hypothesis that slack inherent in the higher cycle serves to absorb the increased delay after TSP events. However, as the cycle continues to increase, the general increase in delay due to increased cycle length outweighs the benefits of additional slack to absorb truncation. Thus, TSP is less disruptive to side street traffic at moderately higher cycle lengths than the optimal cycle length. This becomes a tradeoff between accepting higher base line delays (No TSP) and lower delay increases after a TSP event. The decision may further be guided by the frequency of TSP requests. For
306

higher v/c ratios (i.e., 1.0) these trends become move severe, with side street delays during TSP at the lower cycle significantly higher than the higher cycle. This result may lead to a decision to not implement TSP under high v/c conditions.

(a)

(b)

Figure 98. Graphs. Cross street delay vs cycle for (a) SBL and (b) SBT

Cross street Delay Extent Figure 99 presents the extent of the change in delay for SBT movement after truncation with a cycle length of 110 seconds. Figure 99 (a) includes the data for any truncation amount up to 10 seconds while (b) only includes the vehicles affected by the maximum truncation of 10 seconds. As discussed earlier the two are very similar as a high proportion of truncations reach maximum truncation for this movement. From the figure, delay increases are experienced up to six to eight cycles after truncation. Similar, although slightly muted, results are seen when the truncation is limited to 5 seconds. For this case the delay increases last for up to four to five cycles.

307

(a)

(b)

Figure 99. Graphs. SBT Delay extent, C= 110s for (a) 0 to 10 second truncation and (b) 10 second truncation only

Figure 100 shows the results for the cycle length of 130 seconds and a maximum truncation of 10 seconds. Delay increases persist for only the first three cycles, and the absolute value of the increase in delay per cycle is also less than in the 110 second cycle scenario. Therefore, moving from a cycle length of 110 seconds to 130 seconds, for the maximum truncation of 10 seconds, the time to dissipate the increase in delay after truncation reduces from up to eight cycles to three cycles. Again, similar trends were seen when the maximum truncation was set to five seconds, with lower overall delays and the impact dissipating in 3 cycles (figure not shown).

308

(a)

(b)

Figure 100. Graphs. SBT Delay extent, C= 130s for (a) 0 to 10 second truncation and (b) 10 second truncation only

Similarly, Figure 101 shows SBT movement delay extent for the 150 second cycle after truncations of up to 10 seconds. Similar results to that of the 130 seconds cycle are seen, again delay increases for approximately three cycles after truncation. A maximum truncation of five seconds demonstrated similar results (figures not shown).

(a)

(b)

Figure 101. Graphs. SBT Delay extent, C= 150s for (a) 0 to 10 second truncation and (b) 10 second truncation only

In summary, increasing the cycle length from the optimum increases the baseline delay,

i.e., the overall intersection delay when there is no TSP. However, the higher cycle

lengths provide more slack time to absorb delays from TSP events. For high v/c ratios the

delay due to TSP may be unacceptable, with the impact potentially lasting many cycles.

While a cycle length increase may result in an overall delay increase, it may also provide

a means to allow for acceptable service during TSP.

309

Impact of travel time variability and ETA selection and setting on TSP performance Bus travel time variability
Figure 102 shows the variability of travel time for the bus from check in detector to the stop line at a v/c ratio of 1.0. Only buses that check in during green and traverse the intersection on the same green, including GE, are included. Figure 102 (a) is the actual ETA CDF, while (b) plots actual ETA vs time in cycle. The priority phase has a split of 45 seconds that includes 5 seconds clearance time. The phase split starts at 0. The timing has floating offsets and thus the priority movement occasionally receives an early return to green, depending on the side street traffic demand. Data points outside the normal split on the extreme right of the figure indicate bus arrivals at check-in on early return to green.
From the data, average actual ETA under these conditions is 30.1, the median is 27.0 seconds, and the 85th percentile is 42.0 seconds. Figure 102 (b) shows the expected trend, i.e., that the buses arriving when the priority phase has just started take longer to reach the stop line as they encounter heavier queuing. Near the end of the phase, several buses approach the intersection at free flow speed, with time estimated between 17 and 20 seconds. Although, even near the end of normal green, there are buses at lower than free flow speed as the queue has not yet cleared, thus setting ETA as free flow time could lead to an ineffective GE.
310

(a) Actual ETA CDF at v/c=1.0

(b) Actual ETA vs green elapsed at v/c =

1.0

Figure 102. Graphs. Bus travel time variability at v/c=1.0.

As shown in Figure 103 (a) and (b), travel time variability at a v/c ratio of 0.95 closely resembles the behavior at a v/c ratio of 1.0, although with a higher proportion of buses at free flow near the end of normal green. For the v/c ratio of 0.95 the actual ETA distribution has an average of 26.0 seconds, a median of 22.0 seconds, and 85th percentile of 38.0 seconds. As shown in Figure 103 (c) and (d) the travel time at 0.85 has a much lower variability, with most of the buses arriving after 20 seconds into green experiencing free flow speed through the intersection. For the v/c ratio of 1.0 approximately 19% of the buses have an actual ETA less than 20seconds, this percentage increases to 40% and 70% for the v/c ratios of 0.95 and 0.85, respectively. Higher congestion levels therefore pose a more significant challenge for selection of a single ETA values.

311

(a) ETA CDF at v/c=0.95

(b) ETA vs green elapsed at v/c = 0.95

(c) ETA CDF at v/c=0.85

(d) ETA vs green elapsed at v/c = 0.85

Figure 103. Graphs. Bus travel time variability at v/c ratios of 0.95 and 0.85

ETA sensitivity 312

TSP performance results for different ETA selections at v/c = 1.0 are shown below. Table 15 shows the TSP strategies selected and related GE effectiveness. From the table it is seen that the number of both GE and EG requests are higher for lower ETA values and decreases progressively as ETA increases. The number of unsuccessful GEs is also higher for lower ETAs and decreases progressively as the ETA increases. At lower ETAs, there are more GE attempts but also more GE failures due to underestimation of the actual bus travel times and consequently the bus failing to pass through the intersection before the expiration of maximum GE. Thus, lower ETAs increase GE opportunities but present a risk of additional unsuccessful GEs, while higher ETAs may have missed GE opportunities.

Table 15. TSP Effectiveness under different ETA values

TSP_strategy ETA

EG

GE

GE_Eff GE_NoEff No_Action

17

138

121

97

24

41

20

129

123

101

22

48

25

119

121

99

22

60

30

115

112

94

18

73

GE&EG

35

109

103

87

16

88

40

104

91

75

16

105

45

96

80

69

11

124

50

94

68

59

9

138

17

0

134

110

24

166

GEonly

20

0

134

112

22

166

313

25

0

126

104

22

174

30

0

114

96

18

186

35

0

104

88

16

196

40

0

92

76

16

208

45

0

92

76

16

208

50

0

69

60

9

231

Figure 104 (a) shows the bus travel time for a GE and EG strategy implementation, while Figure 104 (b) is GE only. In the figure ETA17 stands for an ETA of 17 seconds and so forth. The error bars represent 95th confidence intervals. All buses are included in the data whether they receive TSP or not, as the requests for TSP are partly determined by the ETA values. There is a trend of increasing bus travel time as ETA values increase. This is mainly because fewer GE opportunities are identified. Additionally, lower ETAs encourage more aggressive truncation which can clear the queue ahead of the bus.

(e) GE and EG

(f) GE only 314

Figure 104. Graphs. Impact of ETA selection on bus travel time (All Buses)
Figure 105 (a) and (b) show the cross street SBL and SBT delay for the different ETA selections with the GE only strategy implemented. The lower ETA values show higher delays due to an increase in granted TSP requests as well as a higher likelihood of failed GE.

(a)

(b)

Figure 105. Graphs. Cross street Delay Vs ETA for (a) SBL and (b) SBT

AVL Algorithm Table 16 shows the strategies and effectiveness of TSP implemented with the modified AVL algorithm for the v/c ratio of 1.0. From the GE_NoEff column, comparing AVL and fixed location CI-CO TSP, the failed extensions reduce from 19 to 3 in the GE and EG case and from 19 to 1 in the GE only case. This shows the potential for greatly improving TSP effectiveness using an AVL system incorporating CV data.
315

Table 16. TSP strategies and Effectiveness with AVL Algorithm

TSP_strategy Det

EG

GE

GE_Eff GE_NoEff No_Action

AVL

161

105

102

3

24

GE&EG

CI-CO

153

113

94

19

24

AVL

0

120

119

1

170

GEonly

CI-CO

0

131

112

19

159

Figure 106 shows a comparison of the AVL TSP system and the fixed location detector CI-CO for both GE & EG and GE only strategies. Only buses that receive TSP are included in this data. Whereas there is minimal difference between the two systems in the GE and EG implementation, AVL performs better in the GE only implementation. The gain is derived from the reduced green failures. Larger benefits may be expected at higher v/c ratios and for the nearside bus stops where the ETA uncertainty is increased by the dwell time at the bus stop. A follow up study will formulate an AVL algorithm to implement TSP at nearside bus stops including updating priority requests depending on whether the bus stops or skips the bus stop.

316

GE and EG

GE only

Figure 106. Graphs. Bus travel time, Comparing AVL and fixed location check-in.

Impact of dwell time magnitudes and variability Dwell time magnitudes and variability at nearside bus stop
As can be expected the dwell time magnitude at a far side bus stop affects the arrival profile of the bus at the check-in detector. Figure 107 shows the number of buses arriving at the check-in detector at different points in the cycle with (a) plotting data for DT01, (b) for DT02, and (c) for DT03 (see Figure 81), i.e., increasing likelihood and length of stopping profiles. The red vertical line is drawn through the force off point for the priority movement in which the bus is flowing. Moving from DT01 through DT03, an increasing number of buses arrive at the detector after the force off point (to the right of the red line).

(a) Bus arrival profile at check-in DT01 (b) Bus arrival profile at check-in DT01 317

(c) Bus arrival profile at check-in DT03
Figure 107. Graphs. Bus arrival profile at check-in
In the GE only strategy the buses arriving past the force off point can only request EG or must wait for the next programmed green, as their actual ETA will exceed the maximum allowable. This is illustrated in Table 17. The number of buses calling for EG increases progressively from DT01 to DT03 while the number requesting GE decreases.

Table 17. TSP Effectiveness strategies for different dwell time magnitudes

Treat

DT

EG

GE GE_Eff GE_NoEff No_Action

DT01

78

127

118

9

95

GE&EG DT02 117

104

97

7

79

DT03 153

84

81

3

63

DT01

0

124

116

8

176

GEonly DT02

0

105

96

9

195

DT03

0

87

83

4

213

Figure 108 shows bus travel time for the three dwell time distributions. Figure 108 (a) is

for the GE and EG strategy while (b) reflects GE only strategy. Travel time is measured 318

from the check-in detector to the check-out detector immediately past the stop line. This range does not include the far side bus stop; thus, the travel time does not include dwell time and the associated deceleration and acceleration lost times. The plotted data includes all buses regardless of the presence of a TSP request. Despite the high variability in the plots, it is seen that TSP is more beneficial for the lower dwell time magnitudes and variability, such as that in DT01. This is due to the lower dwell times allowing more GE opportunities as more buses can arrive at the detector and request GE prior to the force off point.

(a)

(b)

Figure 108. Graphs. Bus travel time for different dwell time distributions for (a) GE & EG and (b)

EG only

However, the above result is dependent on the offset and indicates that for a given dwell time distribution at the bus stop, signal timing offsets could be fine-tuned to provide more GE opportunities and thus improve TSP performance. Figure 109 shows bus arrival profile results from adjusting the offset from 22 seconds (optimal without consideration

319

of TSP) to 30 and 40 seconds for DT03. The resulting proportions of EG and GE are shown in Table 18.

(a) Offset 22

(b) Offset 30

(c) Offset 40
Figure 109. Graphs. Bus arrival profile at check-in for DT03 and different offsets

Table 18. TSP Strategies for DT03 with different offsets

offset EG

GE

GE_Eff GE_NoEff No_Action

22

153

84

81

3

63

320

30

122

99

97

2

79

40

86

113 112

1

101

As expected, increasing the offsets from 22 to 30 and 40 seconds increased the share of GE as more buses were able to check-in before the signal passed the priority movement force off point. This results in improved average bus travel times as shown in Figure 110. In the figure all buses are included whether they receive TSP or not. Although, as the optimal offsets for buses may not be optimal for the non-transit vehicles this balance needs to be evaluated considering the traffic demand, bus headways, and bus and vehicle occupancies, among other factors.

Figure 110. Graph. Bus Travel Time for DT03 and different offsets
Comparing TSP performance for far side bus stops and nearside bus stops. Table 19 summaries the TSP strategies and effectiveness for far and nearside bus stops for the three dwell time distributions. The most informative column is "GE_NoEff,"
321

which represents the failed green extensions. The column "% GE_NoEff" represents the percentage of the failed GE compared to the total GE. For the nearside bus stops, GE ineffectiveness increases as the magnitudes and variability of dwell time increases. Buses with a higher value of dwell time than used in the estimate of ETA may be granted GE but fail to make it through the intersection.

Table 19.TSP strategies and effectiveness at far and nearside bus stops

stop

TSP

%

No

location Strategy DT EG GE GE_Eff GE_NoEff GE_NoEff Action

DT01 75 129 121

8

GE &

DT02 117 104 97

7

EG

DT03 153 84

81

3

Farside

DT01 0 126 119

7

6.2

96

6.7

79

3.6

63

5.6

174

GE only DT02 0 105 96

9

8.6

195

DT03 0 87

83

4

4.6

213

DT01 61 143 104

39

GE &

DT02 80 146 94

52

EG

DT03 98 156 74

82

Nearside

DT01 0 146 108

38

27.3

96

35.6

74

52.6

46

26.0

154

GE only DT02 0 152 104

48

31.6

148

DT03 0 166 81

85

51.2

134

322

Figure 111 presents a comparison of bus travel time for far side and nearside systems. In both cases, travel time is measured for the same distance starting upstream of the far side bus stop location. Thus, travel time includes dwell time at the bus stop in both cases. Additionally, only buses receiving GE are included in the plots. Comparing the TSP and No TSP cases, it is shown that (1) TSP saves the bus more time for far side bus stops compared to near side bus stops, (2) the performance of TSP for the near side is much more variable, and (3) for nearside bus stops, the performance of TSP in saving bus travel time deteriorates as the magnitude and variability of dwell time increases.

(a) Dwell time distribution, DT01

(b) Dwell time distribution, DT02

323

(c) Dwell time distribution, DT03
Figure 111. Graph. Bus Travel time for far and nearside bus stops
Considering the ineffectiveness of TSP at nearside bus stops in Table 19, green time is wasted at the expense of side street movements at a much higher rate than with far side bus stops. Figure 112 shows a comparison of cross street SBL delay for nearside and far side stops and the three dwell time distributions. The plots show cross street passenger car delays in the four cycles after the bus check-in, whether the bus receives TSP or not. It is noted that (1) delay caused by TSP is higher for nearside bus stops compared to far side stops and (2) for nearside bus stops the delay caused by TSP increases with dwell time magnitudes and variability.
324

Figure 112. Graph. Cross street SBL Delay, Farside vs Nearside bus stop
Conclusions Transit Signal Priority (TSP) aims to provide transit vehicles a free flow path through the intersection, or at least reduce wait time. In the literature the effectiveness of TSP is mixed, having been shown to reduce transit vehicle delays, improve reliability and schedule adherence, and mitigate bus bunching in some instances, while having minimal effect in others. TSP systems are typically designed for corridor operations and involve balancing tradeoffs between providing green time to the priority movement, maintaining arterial coordination, and minimizing the impact to conflicting vehicle's level of service. TSP performance is affected by a wide range of parameters and conditions, including congestion levels, bus headways, bus stop location, detector location, green extension
325

limit, bus arrival time within the cycle, bus stop dwell time, TSP strategy selected, and uncertainty in the arrival time prediction, that is, the Estimated Time of Arrival (ETA),
This study uses a VISSIM simulation environment to evaluate the performance of TSP strategies and establish the critical factors and conditions that affect TSP performance. Critical items assessed are TSP strategies, stop dwell time, ETA, general traffic demand, signal timing parameters, bus stop location (near side vs far side), and TSP triggering strategies both with check-in-check-out (CI-CO) detectors and AVL with CV data.
As such, through a series of simulation experiments the following findings were determined:
8) Several key findings when considering Green Extension (GE), i.e., extending the current green phase to allow a bus to traverse the intersection, and Early Green (EG) i.e., truncating the red to allow a stopped (or arriving) bus to depart sooner, include: a. GE tends to provide more significant bus travel time improvements than EG. This finding is intuitively rational, as with GE the bus skips the entire red, whereas with EG the bus travel time is only reduced by the amount of red truncation. b. GE tends to have a lesser impact on conflicting movements than EG. The delay for the same number of seconds of extension versus truncation tends to be higher and last for more cycles when truncation (i.e., EG) is granted. This is primarily a result of the ability with GE to return to the coordinated timing plan by proportionally shortening all phases in the next cycle, whereas EG tends to require
326

all phase time reduction to come from only the truncated phases preceding the TSP movement. 9) As congestion increases (modeled by increasing v/c ratios in this study) several findings are noted: a. With increasing congestion, the percentage of buses requesting TSP increases due to slower speeds, increased presence of queuing and side street phases maxing out. b. The likelihood of an ineffective GE increases with congestion, i.e., in an ineffective GE a bus requests TSP and GE is provided; however, maximum GE is reached before the bus traverses the intersection and the bus is stopped. c. When providing both GE and EG the bus travel time is lower at lower v/c ratios. This is primarily a result of increasing percentages of EG as v/c increases, where the benefits of EG is less as previous noted. Also, the percentage of inefficient GEs increases with increasing congestion. d. However, when providing both GE and EG the absolute travel time improvement between TSP and non-TSP is greater at higher v/c's, as the non-TSP travel time increases more rapidly than the TSP travel time as v/c increases. e. Similar trends are seen in the travel time of buses that receive TSP when only GE is utilized; however, the improvements are more significant, further demonstrating the advantage of GE. f. The impact to side street delay increases as v/c increases. This is a result of the reduced flexibility to recover as demand increases for the green time.
327

g. As v/c increases the number of cycles for the sides side street delays to return to non-TSP levels increases. For v/c ratios of 1.0 the side street may not return to non-TSP levels for an extended number of cycles past the TSP event, at least 14 cycles in this study. Thus, the likelihood of the impact of the previous bus still being present when the next bus arrives increases as v/c increases.
10) As cycle length increases (for the same volume set): a. The proportion of buses requesting TSP decreases. b. The proportion of buses requesting GE decreases. c. The effectiveness of GE requests improves. d. Bus travel time averaged across all buses increases slightly. e. Cross street delay due to TSP is lowest when the cycle is moderately higher than optimal (calculated to optimize non-transit) movements. f. Overall cross street delay increases as cycle length increases beyond optimal. g. The number of cycles required to dissipate the effect to the side street reduces as well as the impact per cycle.
11) The following is the impact of travel time variability and ETA selection, where only a single ETA may be entered into the traffic controller: a. A short ETA may result in higher GE ineffectiveness, where the maximum GE is exceeded before the bus traverses the intersection. b. A long ETA may result in GE not being granted in situations where the bus would have been able to traverse the intersection.
328

c. As v/c increases travel time variability increases, making the selection of ETA more challenging. Lower v/c results in a higher percentage of buses at free flow speed and improves the effectives of the lower ETA.
d. Overall average bus travel time tends to be lower at the lower ETA. e. Overall impact on the cross street tends to be lower at the higher ETA. 12) When using AVL TSP effectiveness may be improved by mitigating some of the tradeoff that occurs between the high and low ETA values. 13) Dwell time can have a significant impact on TSP, introducing significant uncertainty in the bus arrivals, thus reducing the effectiveness of TSP. Setting of offsets to account for dwell time uncertainty may help improve TSP performance. 14) Near side bus stops introduce more uncertainty in bus arrival time, particularly highlighting the need for an AVL based TSP solution that allows for real-time ETA updates.
Overall, it is seen that TSP performance is most favorable in lower v/c conditions where far side bus stops are present. The lower v/c and greater distance between the upstream bus stop and the intersection lessens the uncertainty in ETA, which is critical to TSP effectiveness. Green extension also provides the most benefit to individual buses. However, as congestion increases the effectiveness of TSP decreases. On a highly congested corridor, i.e., v/c ratios approaching or exceeding 1.0, it is possible that TSP may become infeasible as the non-TSP movements may have insufficient slack in available capacity to recover from the TSP related green truncation.
329

Underlying TSP is a balancing of transit and other vehicle demands. The signal timing for optimal transit vehicle performance may not be optimal for other vehicles in the network. For instance, it was seen that slightly higher cycle lengths or adjusted offsets compared to those for demand based optimal signal settings may result in better TSP performance, as well as lower impacts to non-transit vehicles during TSP events. However, during cycles without TSP, the delay experienced by the non-transit vehicles will be higher than under the lower cycle scenario. One implication of this finding is that the setting of a corridor's signal timing parameters should reflect the corridor purpose. That is, if the corridor is designated to serve a significant transit function, then base timing parameters should be selected to improve bus travel time and to lessen TSP impacts on general traffic. Where transit is not a primary focus then base signal timings such as higher cycle lengths or TSP based offset may not be warranted.
A drawback of this effort is that the study is currently limited to one or two coordinated intersections and a limited number of demand sets. While this study allowed for an exploration of fundamental TSP principles, a next step should involve testing on a realworld corridor using existing conditions and bus schedules. Such as study should include stepping from a simulated analysis, to software-in-the-loop, to field implementation and evaluation. Finally, it was seen that ETA is a critical parameter in the setting of TSP, particularly in congested conditions. Unfortunately, the controllers considered only allowed for a single ETA parameter. This creates a significant constraint on TSP performance. It was seen that AVL provides a significant opportunity to mitigate this
330

drawback. Future efforts should seek to further expand the AVL algorithm explored and
consider test implementations.
REFERENCES for Appendix D
Al-Sahili, K. A., & Taylor, W. C. (1996). Evaluation of Bus Priority Signal Strategies in Ann Arbor, Michigan. Transportation Research Record, 1554(1), 74-79. https://doi.org/10.1177/0361198196155400110
Anderson, P., & Daganzo, C. (2019). Effect of Transit Signal Priority on Bus Service Reliability. Transportation Research Procedia, 38, 2-19. https://doi.org/10.1016/j.trpro.2019.05.002
Balke, K. N., Dudek, C. L., & Urbanik, T. (2000). Development and Evaluation of Intelligent Bus Priority Concept. Transportation Research Record, 1727(1), 1219. https://doi.org/10.3141/1727-02
Beak, B., Zamanipour, M., Head, K. L., & Leonard, B. (2018). Peer-to-Peer Priority Signal Control Strategy in a Connected Vehicle Environment. Transportation Research Record, 2672(18), 15-26. https://doi.org/10.1177/0361198118773567
Bhaskar, A., Chung, E., de Mouzon, O., & Dumont, A.-G. (2007). Methodology for Travel Time Estimation on a Signalised Arterial.
Board, T. R., National Academies of Sciences, E., & Medicine. (2013). Transit Capacity and Quality of Service Manual, Third Edition. The National Academies Press. https://doi.org/doi:10.17226/24766
Board, T. R., National Academies of Sciences, E., & Medicine. (2020). Transit Signal Priority: Current State of the Practice. The National Academies Press. https://doi.org/doi:10.17226/25816
Cvijovic, Z., Zlatkovic, M., Stevanovic, A., & Song, Y. (2022). Conditional Transit Signal Priority for Connected Transit Vehicles. Transportation Research Record, 2676(2), 490-503. https://doi.org/10.1177/03611981211044459
Dai, G., Wang, H., & Wang, W. (2015). Signal optimization and coordination for bus progression based on MAXBAND. KSCE Journal of Civil Engineering. https://doi.org/10.1007/s12205-015-1516-4
331

Ding, J., Yang, M., Wang, W., Xu, C., & Bao, Y. (2015). Strategy for Multiobjective Transit Signal Priority with Prediction of Bus Dwell Time at Stops. Transportation Research Record, 2488(1), 10-19. https://doi.org/10.3141/2488-02
Dion, F., Abu-Lebdeh, G., & Ghanim, M. (2014). The impact of dwell time variability on transit signal priority performance. Canadian Journal of Civil Engineering, 41. https://doi.org/10.1139/cjce-2012-0306
Dion, F., Rakha, H., & Zhang, Y. (2004). Evaluation of Potential Transit Signal Priority Benefits along a Fixed-Time Signalized Arterial. Journal of Transportation Engineering, 130(3), 294-303. https://doi.org/doi:10.1061/(ASCE)0733947X(2004)130:3(294)
Ekeila, W., Sayed, T., & Esawey, M. E. (2009). Development of Dynamic Transit Signal Priority Strategy. Transportation Research Record, 2111(1), 1-9. https://doi.org/10.3141/2111-01
Farid, Y. Z., Christofa, E., & Paget-Seekins, L. (2016). Estimation of Short-Term Bus Travel Time by Using Low-Resolution Automated Vehicle Location Data. Transportation Research Record, 2539(1), 113-118. https://doi.org/10.3141/253913
Han, Y., Kim, M., & Kim, Y. (2022). Progression Control Model to Enhance Performance of Transit Signal Priority. IEEE Access, 10, 14397-14408. https://doi.org/10.1109/ACCESS.2022.3146716
Hu, J., Park, B., & Lee, Y.-J. (2015). Coordinated transit signal priority supporting transit progression under Connected Vehicle Technology. Transportation research part C: emerging technologies, 55. https://doi.org/10.1016/j.trc.2014.12.005
Hu, J., Park, B., & Parkany, A. E. (2014). Transit Signal Priority with Connected Vehicle Technology. Transportation Research Record, 2418(1), 20-29. https://doi.org/10.3141/2418-03
Hu, J., Park, B. B., & Lee, Y.-J. (2016). Transit signal priority accommodating conflicting requests under Connected Vehicles technology. Transportation research part C: emerging technologies, 69, 173-192. https://doi.org/https://doi.org/10.1016/j.trc.2016.06.001
Isukapati, I. K., Rudov, H., Barlow, G. J., & Smith, S. F. (2017). Analysis of Trends in Data on Transit Bus Dwell Times. Transportation Research Record, 2619(1), 6474. https://doi.org/10.3141/2619-07
332

Jeong, Y., & Kim, Y. (2014). Tram passive signal priority strategy based on the MAXBAND model. KSCE Journal of Civil Engineering, 18(5), 1518-1527. https://doi.org/10.1007/s12205-014-0159-1
Jian, D., Lu, Z., & Yan, Z. (2013, 20-23 June 2013). Mixed model for prediction of bus arrival times. 2013 IEEE Congress on Evolutionary Computation,
Kim, W., & Rilett, L. R. (2005). Improved Transit Signal Priority System for Networks with Nearside Bus Stops. Transportation Research Record, 1925(1), 205-214. https://doi.org/10.1177/0361198105192500121
Koonce, P., Ringert, J., & Urbanik, T. (2002). Detection Range Setting Methodology for Signal Priority. Journal of Public Transportation, 5. https://doi.org/10.5038/23750901.5.2.6
Kumar, B. A., Jairam, R., Arkatkar, S. S., & Vanajakshi, L. (2019). Real time bus travel time prediction using k-NN classifier. Transportation Letters, 11(7), 362-372. https://doi.org/10.1080/19427867.2017.1366120
Lee, W.-H., & Wang, H.-C. (2022). A Person-Based Adaptive Traffic Signal Control Method with Cooperative Transit Signal Priority. Journal of advanced transportation, 2022, 2205292. https://doi.org/10.1155/2022/2205292
Lee, Y.-J., Dadvar, S., Hu, J., & Park, B. B. (2017). Transit Signal Priority Experiment in a Connected Vehicle Technology Environment. Journal of Transportation Engineering, Part A: Systems, 143(8), 05017005. https://doi.org/doi:10.1061/JTEPBS.0000062
Li, M., Yin, Y., Zhang, W.-B., Zhou, K., & Nakamura, H. (2011). Modeling and Implementation of Adaptive Transit Signal Priority on Actuated Control Systems. Computer-Aided Civil and Infrastructure Engineering, 26(4), 270-284. https://doi.org/https://doi.org/10.1111/j.1467-8667.2010.00677.x
Li, Y., Li, M., Li, Y., Beaird, S., Skabardonis, A., Koonce, P., Zhou, K., Zhang, W.-B., Hegen, L., Hu, K., & Sun, Z. S. (2008). Transit signal priority research tools [Tech Report]. https://rosap.ntl.bts.gov/view/dot/16521
Liang, Y., Wu, Z., Li, J., Li, F., & Wang, Y. (2018). Shockwave-Based Queue Length Estimation Method for Presignals for Bus Priority. Journal of Transportation Engineering, Part A: Systems, 144(9), 04018057. https://doi.org/doi:10.1061/JTEPBS.0000175
333

Liao, C.-F., & Davis, G. A. (2007). Simulation Study of Bus Signal Priority Strategy:Taking Advantage of Global Positioning System, Automated Vehicle Location System, and Wireless Communications. Transportation Research Record, 2034(1), 82-91. https://doi.org/10.3141/2034-10
Little, J. D., Kelson, M. D., & Gartner, N. H. (1981). MAXBAND: A versatile program for setting signals on arteries and triangular networks.
Liu, H., Lin, W.-H., & Tan, C.-w. (2007). Operational Strategy for Advanced Vehicle Location System&#x2013;Based Transit Signal Priority. Journal of Transportation Engineering, 133(9), 513-522. https://doi.org/doi:10.1061/(ASCE)0733-947X(2007)133:9(513)
Liu, H. X., Wu, X., Ma, W., & Hu, H. (2009). Real-time queue length estimation for congested signalized intersections. Transportation research part C: emerging technologies, 17(4), 412-427. https://doi.org/https://doi.org/10.1016/j.trc.2009.02.003
Mishra, S., Kattan, L., & Wirasinghe, S. C. (2020). Transit Signal Priority Along a Signalized Arterial: A Passenger-based Approach. ACM Trans. Spatial Algorithms Syst., 6(1), Article 5. https://doi.org/10.1145/3355611
Mohammadi, R., Roncoli, C., & Mladenovic, M. N. (2020, 3-5 Nov. 2020). Transit signal priority in a connected vehicle environment: User throughput and schedule delay optimization approach. 2020 Forum on Integrated and Sustainable Transportation Systems (FISTS),
Muthuswamy, S., McShane, W. R., & Daniel, J. R. (2007). Evaluation of Transit Signal Priority and Optimal Signal Timing Plans in Transit and Traffic Operations. Transportation Research Record, 2034(1), 92-102. https://doi.org/10.3141/203411
National Academies of Sciences Engineering and Medicine. (2020). Transit Signal Priority: Current State of the Practice. https://nap.nationalacademies.org/catalog/25816/transit-signal-priority-currentstate-of-the-practice
Ngan, V. (2004). Impacts of Various Traffic Parameters on Transit Signal Priority Effectiveness Impacts of Various Parameters on Transit Signal Priority Effectiveness [Text]. Journal of Public Transportation, 7(3).
334

NTCIP Signal Control Priority Working Group., American Association of State Highway and Transportation Officials., Institute of Transportation Engineers., National Electrical Manufacturers Association., & Joint Committee on the NTCIP. (2014). National Transportation Communications for ITS Protocol : objects definitions for signal control and prioritization. AASHTO : ITE ;
NEMA.
PTV. (2020). Ring Barrier Controller User Manual PTV AG.
Rakha, H., & Zhang, Y. (2004). Sensitivity Analysis of Transit Signal Priority Impacts on Operation of a Signalized Intersection. Journal of Transportation Engineering, 130(6), 796-804. https://doi.org/doi:10.1061/(ASCE)0733-947X(2004)130:6(796)
Shalaby, A., & Farhan, A. (2004). Prediction Model of Bus Arrival and Departure Times Using AVL and APC Data. Journal of Public Transportation, 7. https://doi.org/10.5038/2375-0901.7.1.3
Sheffield, M. H., Schultz, G. G., Bassett, D., & Eggett, D. L. (2021). Sensitivity Analysis of the Transit Signal Priority Requesting Threshold and the Impact on Bus Performance and General Traffic. Transportation Research Record, 2675(5), 149163. https://doi.org/10.1177/0361198120985853
Smith, H. R., Hemily, P. B., Ivanovic, M., ITS America., & United States. Department of Transportation. (2005). Transit signal priority (TSP) : a planning and implementation handbook. http://www.fta.dot.gov/documents/TSPHandbook1020-05.pdf
Tan, C.-W., Park, S., Liu, H., Xu, Q., & Lau, P. (2008). Prediction of Transit Vehicle Arrival Time for Signal Priority Control: Algorithm and Performance. IEEE Transactions on Intelligent Transportation Systems, 9, 688-696.
Taparia, A., & Brady, M. (2021, 16-17 June 2021). Bus Journey and Arrival Time Prediction based on Archived AVL/GPS data using Machine Learning. 2021 7th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS),
Teng, K., Liu, H., & Rai, L. (2019). Transit Priority Signal Control Scheme Considering the Coordinated Phase for Single-Ring Sequential Phasing Under Connected Vehicle Environment. IEEE Access, 7, 61057-61069. https://doi.org/10.1109/ACCESS.2019.2915665
335

Tu, T., Sano, K., & Y, N. (2012). BUS PRIORITY STRATEGY COMPARISONS AT SIGNALIZED INTERSECTIONS. Journal of Japan Society of Civil Engineers, Ser. D3 (Infrastructure Planning and Management), 68, I_719-I_729. https://doi.org/10.2208/jscejipm.68.I_719
Varaiya, P. (2013). Max pressure control of a network of signalized intersections. Transportation research part C: emerging technologies, 36, 177-195. https://doi.org/https://doi.org/10.1016/j.trc.2013.08.014
Wang, Q., Yang, X., Leonard, B. D., & Mackey, J. (2020). Field Evaluation of Connected Vehicle-Based Transit Signal Priority Control under Two Different Signal Plans. Transportation Research Record, 2674(7), 172-180. https://doi.org/10.1177/0361198120921161
Wu, K., & Guler, S. I. (2019). Estimating the impacts of transit signal priority on intersection operations: A moving bottleneck approach. Transportation research part C: emerging technologies, 105, 346-358. https://doi.org/https://doi.org/10.1016/j.trc.2019.06.003
Xu, H., & Ying, J. (2017). Bus arrival time prediction with real-time and historic data. Cluster Computing, 20(4), 3099-3106. https://doi.org/10.1007/s10586-017-1006-1
Xu, T., Barman, S., Levin, M. W., Chen, R., & Li, T. (2022). Integrating public transit signal priority into max-pressure signal control: Methodology and simulation study on a downtown network. Transportation research part C: emerging technologies, 138, 103614. https://doi.org/https://doi.org/10.1016/j.trc.2022.103614
Yang, K., Menendez, M., & Guler, S. I. (2019). Implementing transit signal priority in a connected vehicle environment with and without bus stops. Transportmetrica B: Transport Dynamics, 7(1), 423-445. https://doi.org/10.1080/21680566.2018.1434019
Zeng, X., Sun, X., Zhang, Y., & Quadrifoglio, L. (2015). Person-Based Adaptive Priority Signal Control with Connected-Vehicle Information. Transportation Research Record, 2487(1), 78-87. https://doi.org/10.3141/2487-07
Zhou, G., Gan, A., & Zhu, X. (2006). Determination of Optimal Detector Location for Transit Signal Priority with Queue Jumper Lanes. Transportation Research Record, 1978(1), 123-129. https://doi.org/10.1177/0361198106197800116
336

337

REFERENCES
(RITIS), R. I. T. I. S. (2021). Regional Integrated Transportation Information System (RITIS). Retrieved July 28 from https://www.ritis.org/traffic/
Al-Sahili, K. A., & Taylor, W. C. (1996). Evaluation of Bus Priority Signal Strategies in Ann Arbor, Michigan. Transportation Research Record, 1554(1), 74-79. https://doi.org/10.1177/0361198196155400110
Allibhai, J. (2018). Hold-out vs. Cross-validation in Machine Learning. https://medium.com/@jaz1/holdout-vs-cross-validation-in-machine-learning7637112d3f8f#:~:text=Cross%2Dvalidation%20is%20usually%20the,just%20one %20train%2Dtest%20split
Anderson, J., Suh, W., Guin, A., Hunter, M., & Rodgers, M. O. (2019). Accounting for composite travel time distributions within a traffic stream in determining Levelof-Service. Journal of Intelligent & Fuzzy Systems, 36(2), 955-965.
Anderson, P., & Daganzo, C. (2019). Effect of Transit Signal Priority on Bus Service Reliability. Transportation Research Procedia, 38, 2-19. https://doi.org/10.1016/j.trpro.2019.05.002
Balke, K. N., Dudek, C. L., & Urbanik, T. (2000). Development and Evaluation of Intelligent Bus Priority Concept. Transportation Research Record, 1727(1), 1219. https://doi.org/10.3141/1727-02
Beak, B., Zamanipour, M., Head, K. L., & Leonard, B. (2018). Peer-to-Peer Priority Signal Control Strategy in a Connected Vehicle Environment. Transportation Research Record, 2672(18), 15-26. https://doi.org/10.1177/0361198118773567
Bhaskar, A., Chung, E., de Mouzon, O., & Dumont, A.-G. (2007). Methodology for Travel Time Estimation on a Signalised Arterial.
Board, T. R., National Academies of Sciences, E., & Medicine. (2013). Transit Capacity and Quality of Service Manual, Third Edition. The National Academies Press. https://doi.org/doi:10.17226/24766
338

Board, T. R., National Academies of Sciences, E., & Medicine. (2020). Transit Signal Priority: Current State of the Practice. The National Academies Press. https://doi.org/doi:10.17226/25816
Buchenscheit, A., Schaub, F., Kargl, F., & Weber, M. (2009, 28-30 Oct. 2009). A VANET-based emergency vehicle warning system. 2009 IEEE Vehicular Networking Conference (VNC),
Chang, E., & Messer, C. J. (1991). Arterial signal timing optimization using PASSER II90-Program user's manual.
Chen, Z., & Huang, X. (2017). End-to-end learning for lane keeping of self-driving cars. 2017 IEEE Intelligent Vehicles Symposium (IV),
Cohen, S. (1982). Concurrent use of the MAXBAND and TRANSYT signal timing programs for arterial signal optimization. Federal Highway Administration, Urban Transportation Management Division.
Comert, G., & Cetin, M. (2021). Queue length estimation from connected vehicles with range measurement sensors at traffic signals. Applied Mathematical Modelling, 99, 418-434.
Cong Minh Ho Chi, C., Matsumoto, S., Sano, K., Cong Minh Doctoral Student, C., Matsumoto Professor, S., & Sano Associate Professor, K. (2005). The speed, flow and headway analyses of motorcycle traffic Title: Two-player Game Theory Based Analysis of Motorcycle Driver's Behavior At Signalized Intersection View project THE SPEED, FLOW AND HEADWAY ANALYSES OF MOTORCYCLE TRAFFIC. https://doi.org/10.11175/easts.6.1496
Corts, C. E., & Stefoni, B. (2023). Trajectory Simulation of Emergency Vehicles and Interactions with Surrounding Traffic. Journal of advanced transportation, 2023. https://doi.org/10.1155/2023/5995950
Cvijovic, Z., Zlatkovic, M., Stevanovic, A., & Song, Y. (2022). Conditional Transit Signal Priority for Connected Transit Vehicles. Transportation Research Record, 2676(2), 490-503. https://doi.org/10.1177/03611981211044459
Dai, G., Wang, H., & Wang, W. (2015). Signal optimization and coordination for bus progression based on MAXBAND. KSCE Journal of Civil Engineering. https://doi.org/10.1007/s12205-015-1516-4
339

Ding, J., Yang, M., Wang, W., Xu, C., & Bao, Y. (2015). Strategy for Multiobjective Transit Signal Priority with Prediction of Bus Dwell Time at Stops. Transportation Research Record, 2488(1), 10-19. https://doi.org/10.3141/2488-02
Dion, F., Abu-Lebdeh, G., & Ghanim, M. (2014). The impact of dwell time variability on transit signal priority performance. Canadian Journal of Civil Engineering, 41. https://doi.org/10.1139/cjce-2012-0306
Dion, F., Rakha, H., & Zhang, Y. (2004). Evaluation of Potential Transit Signal Priority Benefits along a Fixed-Time Signalized Arterial. Journal of Transportation Engineering, 130(3), 294-303. https://doi.org/doi:10.1061/(ASCE)0733947X(2004)130:3(294)
Dowling, R., Skabardonis, A., & Alexiadis, V. (2004). Traffic analysis toolbox, volume III: Guidelines for applying traffic microsimulation modeling software.
Ekeila, W., Sayed, T., & Esawey, M. E. (2009). Development of Dynamic Transit Signal Priority Strategy. Transportation Research Record, 2111(1), 1-9. https://doi.org/10.3141/2111-01
Eliminator. (2021). Preemption System - Emergency Vehicle - Transit. Collison Control Communications. Retrieved July 27 from https://collisioncontrol.net/PreemptionSystems/?gclid=Cj0KCQiApt_xBRDxARI sAAMUMu_IjJEiTe4k0tnTcFLBTB6GNLLJ1F9WstFXn0ty2tgO5SPZwVc1syoa AqMyEALw_wcB
Eltayeb, A. S., Almubarak, H. O., & Attia, T. A. (2013). A GPS based traffic light preemption control system for emergency vehicles. 2013 international conference on computing, electrical and electronic engineering (icceee),
Emtrac. EMTRAC Systems Signal Priority: Emergency Vehicle Preemption (EVP) Utilizing ITS Functionality. Retrieved July 27 from https://www.emtracsystems.com/first-response/emergency-vehicle-preemptionevp/
Farid, Y. Z., Christofa, E., & Paget-Seekins, L. (2016). Estimation of Short-Term Bus Travel Time by Using Low-Resolution Automated Vehicle Location Data. Transportation Research Record, 2539(1), 113-118. https://doi.org/10.3141/253913
Galarnyk, M. (2018). Understanding Boxplots. towards data science. Retrieved July 30 from https://towardsdatascience.com/understanding-boxplots-5e2df7bcbd51
340

Gedawy, H. K., Dias, M., & Harras, K. (2008). Dynamic path planning and traffic light coordination for emergency vehicle routing. Comp. Sci. Dept., Carnegie Mellon Univ., Pittsburgh, Pennsylvania, USA, 189.
Google. (2021). Google Maps: Peachtree Industrial Boulevard. Retrieved July 27 from https://www.google.com/maps/dir/33.9459033,-84.2363882/34.0064816,84.1701493/@33.9949555,-84.2917807,11679m/data=!3m1!1e3!4m2!4m1!3e0
Grzywaczewski, A. (2017). Training ai for self-driving vehicles: the challenge of scale. Available from Internet: https://devblogs. nvidia. com/training-self-drivingvehicles-challenge-scale.
Han, Y., Kim, M., & Kim, Y. (2022). Progression Control Model to Enhance Performance of Transit Signal Priority. IEEE Access, 10, 14397-14408. https://doi.org/10.1109/ACCESS.2022.3146716
Hse, F., Roch, L. M., & Aspuru-Guzik, A. (2019). Next-generation experimentation with self-driving laboratories. Trends in Chemistry, 1(3), 282-291.
Hayasaka, S. (2022, Feb 11, 2022). How Many Clusters? Methods for choosing the right number of clusters. Towards Data Science. Retrieved June 01 from https://towardsdatascience.com/how-many-clusters6b3f220f0ef5#:~:text=For%20some%20clustering%20algorithms%2C%20such,i nformative%20(see%20Figure%201).
Homaei, H., Hejazi, S., & Dehghan, S. A. M. (2015). A new traffic light controller using fuzzy logic for a full single junction involving emergency vehicle preemption. Journal of Uncertain Systems, 9(1), 49-61.
Hong, K.-S., Jung, J.-H., & Ahn, G.-H. (2012). Development of the Emergency Vehicle Preemption Control System Based on UTIS. The Journal of The Korea Institute of Intelligent Transport Systems, 11(2), 39-47.
Hossain, S., Fayjie, A. R., Doukhi, O., & Lee, D.-j. (2018). CAIAS simulator: selfdriving vehicle simulator for AI research. International Conference on Intelligent Computing & Optimization,
How a Connected Emergency Vehicle Preemption System Works. (2019). applied information. Retrieved Nov 05 from https://appinfoinc.com/how-preemptionsystem-works/
341

Hu, J., Park, B., & Lee, Y.-J. (2015). Coordinated transit signal priority supporting transit progression under Connected Vehicle Technology. Transportation research part C: emerging technologies, 55. https://doi.org/10.1016/j.trc.2014.12.005
Hu, J., Park, B., & Parkany, A. E. (2014). Transit Signal Priority with Connected Vehicle Technology. Transportation Research Record, 2418(1), 20-29. https://doi.org/10.3141/2418-03
Hu, J., Park, B. B., & Lee, Y.-J. (2016). Transit signal priority accommodating conflicting requests under Connected Vehicles technology. Transportation research part C: emerging technologies, 69, 173-192. https://doi.org/https://doi.org/10.1016/j.trc.2016.06.001
Information, A. Emergency Vehicle Preemption and Priority System: Applied Information. Retrieved July 27 from appinfoinc.com/solutions/preemptionpriority/.
Isukapati, I. K., Rudov, H., Barlow, G. J., & Smith, S. F. (2017). Analysis of Trends in Data on Transit Bus Dwell Times. Transportation Research Record, 2619(1), 6474. https://doi.org/10.3141/2619-07
Jeong, Y., & Kim, Y. (2014). Tram passive signal priority strategy based on the MAXBAND model. KSCE Journal of Civil Engineering, 18(5), 1518-1527. https://doi.org/10.1007/s12205-014-0159-1
Jian, D., Lu, Z., & Yan, Z. (2013, 20-23 June 2013). Mixed model for prediction of bus arrival times. 2013 IEEE Congress on Evolutionary Computation,
Kamble, S., & Kounte, M. R. (2022). A Survey on Emergency Vehicle Preemption Methods Based on Routing and Scheduling. International Journal of Computer Networks And Applications, 9, 60-71. https://doi.org/10.22247/ijcna/2022/211623
Kim, W., & Rilett, L. R. (2005). Improved Transit Signal Priority System for Networks with Nearside Bus Stops. Transportation Research Record, 1925(1), 205-214. https://doi.org/10.1177/0361198105192500121
Koonce, P., Ringert, J., & Urbanik, T. (2002). Detection Range Setting Methodology for Signal Priority. Journal of Public Transportation, 5. https://doi.org/10.5038/23750901.5.2.6
342

Kumar, B. A., Jairam, R., Arkatkar, S. S., & Vanajakshi, L. (2019). Real time bus travel time prediction using k-NN classifier. Transportation Letters, 11(7), 362-372. https://doi.org/10.1080/19427867.2017.1366120
Kwon, E., Kim, S., & Betts, R. (2003). Route-based dynamic preemption of traffic signals for emergency vehicle operations. Transportation Research Board 82nd Annual MeetingTransportation Research Board,
Lee, W.-H., & Wang, H.-C. (2022). A Person-Based Adaptive Traffic Signal Control Method with Cooperative Transit Signal Priority. Journal of advanced transportation, 2022, 2205292. https://doi.org/10.1155/2022/2205292
Lee, Y.-J., Dadvar, S., Hu, J., & Park, B. B. (2017). Transit Signal Priority Experiment in a Connected Vehicle Technology Environment. Journal of Transportation Engineering, Part A: Systems, 143(8), 05017005. https://doi.org/doi:10.1061/JTEPBS.0000062
Li, J.-Q., Zhou, K., Shladover, S. E., & Skabardonis, A. (2013). Estimating queue length under connected vehicle technology: Using probe vehicle, loop detector, and fused data. Transportation Research Record, 2356(1), 17-22.
Li, M., Yin, Y., Zhang, W.-B., Zhou, K., & Nakamura, H. (2011). Modeling and Implementation of Adaptive Transit Signal Priority on Actuated Control Systems. Computer-Aided Civil and Infrastructure Engineering, 26(4), 270-284. https://doi.org/https://doi.org/10.1111/j.1467-8667.2010.00677.x
Li, Y., Li, M., Li, Y., Beaird, S., Skabardonis, A., Koonce, P., Zhou, K., Zhang, W.-B., Hegen, L., Hu, K., & Sun, Z. S. (2008). Transit signal priority research tools [Tech Report]. https://rosap.ntl.bts.gov/view/dot/16521
Liang, Y., Wu, Z., Li, J., Li, F., & Wang, Y. (2018). Shockwave-Based Queue Length Estimation Method for Presignals for Bus Priority. Journal of Transportation Engineering, Part A: Systems, 144(9), 04018057. https://doi.org/doi:10.1061/JTEPBS.0000175
Liao, C.-F., & Davis, G. A. (2007). Simulation Study of Bus Signal Priority Strategy:Taking Advantage of Global Positioning System, Automated Vehicle Location System, and Wireless Communications. Transportation Research Record, 2034(1), 82-91. https://doi.org/10.3141/2034-10
Lidestam, B., Thorslund, B., Selander, H., Nsman, D., & Dahlman, J. (2020). In-Car Warnings of Emergency Vehicles Approaching: Effects on Car Drivers'
343

Propensity to Give Way. Frontiers in Sustainable Cities, 2. https://doi.org/10.3389/frsc.2020.00019
Little, J. D., Kelson, M. D., & Gartner, N. H. (1981). MAXBAND: A versatile program for setting signals on arteries and triangular networks.
Liu, H., Lin, W.-H., & Tan, C.-w. (2007). Operational Strategy for Advanced Vehicle Location System&#x2013;Based Transit Signal Priority. Journal of Transportation Engineering, 133(9), 513-522. https://doi.org/doi:10.1061/(ASCE)0733-947X(2007)133:9(513)
Liu, H. X., Wu, X., Ma, W., & Hu, H. (2009). Real-time queue length estimation for congested signalized intersections. Transportation research part C: emerging technologies, 17(4), 412-427. https://doi.org/https://doi.org/10.1016/j.trc.2009.02.003
McHale, G. M., & Collura, J. (2003). Improving emergency vehicle traffic signal priority system assessment methodologies. Transportation Research Board 82nd Annual MeetingTransportation Research Board,
McKight, P. E., & Najab, J. (2010). Kruskal-wallis test. The corsini encyclopedia of psychology, 1-1.
Min, W., Yu, L., Chen, P., Zhang, M., Liu, Y., & Wang, J. (2019). On-demand greenwave for emergency vehicles in a time-varying road network with uncertainties. IEEE Transactions on Intelligent Transportation Systems, 21(7), 3056-3068.
Mishra, S., Kattan, L., & Wirasinghe, S. C. (2020). Transit Signal Priority Along a Signalized Arterial: A Passenger-based Approach. ACM Trans. Spatial Algorithms Syst., 6(1), Article 5. https://doi.org/10.1145/3355611
Mittal, M. S. (2002). Assessing the Performance of an Emergency Vehicle Preemption System: A Case Study on US 1 in Fairfax County, Virginia Virginia Tech].
Mohammadi, R., Roncoli, C., & Mladenovic, M. N. (2020, 3-5 Nov. 2020). Transit signal priority in a connected vehicle environment: User throughput and schedule delay optimization approach. 2020 Forum on Integrated and Sustainable Transportation Systems (FISTS),
Muthuswamy, S., McShane, W. R., & Daniel, J. R. (2007). Evaluation of Transit Signal Priority and Optimal Signal Timing Plans in Transit and Traffic Operations. 344

Transportation Research Record, 2034(1), 92-102. https://doi.org/10.3141/203411
National Academies of Sciences Engineering and Medicine. (2020). Transit Signal Priority: Current State of the Practice. https://nap.nationalacademies.org/catalog/25816/transit-signal-priority-currentstate-of-the-practice
Nelson, E. J., & Bullock, D. (2000). Impact of emergency vehicle preemption on signalized corridor operation: An evaluation. Transportation Research Record, 1727(1), 1-11.
Ngan, V. (2004). Impacts of Various Traffic Parameters on Transit Signal Priority Effectiveness Impacts of Various Parameters on Transit Signal Priority Effectiveness [Text]. Journal of Public Transportation, 7(3).
Nguyen, Q., Valizadegan, H., & Hauskrecht, M. (2014). Learning classification models with soft-label information. Journal of the American Medical Informatics Association, 21(3), 501-508.
Noori, H., Fu, L., & Shiravi, S. (2016). A connected vehicle based traffic signal control strategy for emergency vehicle preemption. Transportation Research Board 95th Annual Meeting,
NTCIP Signal Control Priority Working Group., American Association of State Highway and Transportation Officials., Institute of Transportation Engineers., National Electrical Manufacturers Association., & Joint Committee on the NTCIP. (2014). National Transportation Communications for ITS Protocol : objects definitions for signal control and prioritization. AASHTO : ITE ;
NEMA.
Obenberger, J., & Collura, J. (2001). Transition strategies to exit preemption control: State-of-the-practice assessment. Transportation Research Record, 1748(1), 7279.
Obrusnk, V., Herman, I., & Hurk, Z. (2020). Queue discharge-based emergency vehicle traffic signal preemption. IFAC-PapersOnLine, 53, 14997-15002. https://doi.org/10.1016/j.ifacol.2020.12.1998
OpenStreetMap. OpenStreetMap: Peachtree Industrial Boulevard Corridor. OpenStreetMap. Retrieved January 17 from 345

https://www.openstreetmap.org/directions?engine=fossgis_osrm_car&route=33.9 4488%2C-84.23857%3B34.01961%2C-84.16084#map=13/33.9823/-84.1997
PTV. (2020). Ring Barrier Controller User Manual PTV AG.
PTV, A. (2014). PTV Vissim 2021: RBC user manual. PTV AG: Karlsruhe, Germany.
PTV, A. (2021a). PTV Vissim- 3D Vehicle Models. PTV AG: Karlsruhe, Germany. Retrieved July 28 from https://company.ptvgroup.com/en-us/ptv-vissim-3dvehicle-models
PTV, A. (2021b). PTV Vissim 2021 user manual. PTV AG: Karlsruhe, Germany.
Python. (2018). Python 3.7.0 Release. Retrieved June 25, 2020 from https://www.python.org/downloads/release/python-370/
Qin, X., & Khan, A. M. (2012). Control strategies of traffic signal timing transition for emergency vehicle preemption. Transportation research part C: emerging technologies, 25, 1-17.
Raj, A. (2020). Unlocking the True Power of Support Vector Regression. Retrieved March 15 from https://towardsdatascience.com/unlocking-the-true-power-ofsupport-vector-regression-847fd123a4a0
Rakha, H., & Zhang, Y. (2004). Sensitivity Analysis of Transit Signal Priority Impacts on Operation of a Signalized Intersection. Journal of Transportation Engineering, 130(6), 796-804. https://doi.org/doi:10.1061/(ASCE)0733-947X(2004)130:6(796)
Rao, Q., & Frtunikj, J. (2018). Deep learning for self-driving cars: Chances and challenges. Proceedings of the 1st International Workshop on Software Engineering for AI in Autonomous Systems,
Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20, 53-65.
Saravanan, R., Paul, P. V., Nivedha, M., & Nirosha, B. (2017). Optimal path planning and traffic signal control system for emergency vehicle--A survey. 2017 International Conference on Computation of Power, Energy Information and Commuincation (ICCPEIC),
346

Savolainen, P. T., Datta, T. K., Ghosh, I., & Gates, T. J. (2010). Effects of dynamically activated emergency vehicle warning sign on driver behavior at Urban intersections. Transportation Research Record(2149). https://doi.org/10.3141/2149-09
Scikit-learn. scikit-learn: sklearn.neural_network.MLPRegressor. Retrieved July 15 from https://scikitlearn.org/stable/modules/generated/sklearn.neural_network.MLPRegressor.html
Scikit-learn. sklearn.ensemble.RandomForestRegressor. Retrieved September 12 from https://scikitlearn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.ht ml
Shaaban, K., Khan, M. A., Hamila, R., & Ghanim, M. (2019a). A strategy for emergency vehicle preemption and route selection. Arabian Journal for Science and Engineering, 44(10), 8905-8913.
Shaaban, K., Khan, M. A., Hamila, R., & Ghanim, M. (2019b). A Strategy for Emergency Vehicle Preemption and Route Selection. Arabian Journal for Science and Engineering, 44(10). https://doi.org/10.1007/s13369-019-03913-8
Shalaby, A., & Farhan, A. (2004). Prediction Model of Bus Arrival and Departure Times Using AVL and APC Data. Journal of Public Transportation, 7. https://doi.org/10.5038/2375-0901.7.1.3
Sheffield, M. H., Schultz, G. G., Bassett, D., & Eggett, D. L. (2021). Sensitivity Analysis of the Transit Signal Priority Requesting Threshold and the Impact on Bus Performance and General Traffic. Transportation Research Record, 2675(5), 149163. https://doi.org/10.1177/0361198120985853
Shoufeng, L., Ximin, L., & Shiqiang, D. (2008). Revised MAXBAND model for bandwidth optimization of traffic flow dispersion. 2008 ISECS International Colloquium on Computing, Communication, Control, and Management,
Shulga, D. (2018). 5 Reasons why you should use Cross-Validation in your Data Science Projects. https://towardsdatascience.com/5-reasons-why-you-should-use-crossvalidation-in-your-data-science-project-8163311a1e79
Smith, H. R., Hemily, P. B., Ivanovic, M., ITS America., & United States. Department of Transportation. (2005). Transit signal priority (TSP) : a planning and
347

implementation handbook. http://www.fta.dot.gov/documents/TSPHandbook1020-05.pdf
Tan, C.-W., Park, S., Liu, H., Xu, Q., & Lau, P. (2008). Prediction of Transit Vehicle Arrival Time for Signal Priority Control: Algorithm and Performance. IEEE Transactions on Intelligent Transportation Systems, 9, 688-696.
Taparia, A., & Brady, M. (2021, 16-17 June 2021). Bus Journey and Arrival Time Prediction based on Archived AVL/GPS data using Machine Learning. 2021 7th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS),
Taylor, C. (2019). Understanding the Interquartile Range in Statistics. ThoughtCo. Retrieved August 1 from https://www.thoughtco.com/what-is-the-interquartilerange-3126245
Technologies, G. T. (2021). Central Management Software (CMS) for Traffic Signal Priority Control. Global Traffic Technologies. Retrieved July 27 from www.gtt.com/project/central-management-software-cms-for-traffic-signalpriority-control/
Ten, H., Qi, Y., Falcocchio, J. C., Kim, K.-B., Patel, R., & Athanailos, E. (2003). Simulation testing of adaptive control, bus priority and emergency vehicle preemption in New York City. Transportation Research Board 82nd Annual MeetingTransportation Research Board,
Teng, K., Liu, H., & Rai, L. (2019). Transit Priority Signal Control Scheme Considering the Coordinated Phase for Single-Ring Sequential Phasing Under Connected Vehicle Environment. IEEE Access, 7, 61057-61069. https://doi.org/10.1109/ACCESS.2019.2915665
Tiaprasert, K., Zhang, Y., Wang, X. B., & Zeng, X. (2015). Queue length estimation using connected vehicle technology for adaptive signal control. IEEE Transactions on Intelligent Transportation Systems, 16(4), 2129-2140.
Trucano, T. G., Swiler, L. P., Igusa, T., Oberkampf, W. L., & Pilch, M. (2006). Calibration, validation, and sensitivity analysis: What's what. Reliability Engineering & System Safety, 91(10-11), 1331-1357.
Tu, T., Sano, K., & Y, N. (2012). BUS PRIORITY STRATEGY COMPARISONS AT SIGNALIZED INTERSECTIONS. Journal of Japan Society of Civil Engineers,
348

Ser. D3 (Infrastructure Planning and Management), 68, I_719-I_729. https://doi.org/10.2208/jscejipm.68.I_719
Varaiya, P. (2013). Max pressure control of a network of signalized intersections. Transportation research part C: emerging technologies, 36, 177-195. https://doi.org/https://doi.org/10.1016/j.trc.2013.08.014
Wang, Q., Yang, X., Leonard, B. D., & Mackey, J. (2020). Field Evaluation of Connected Vehicle-Based Transit Signal Priority Control under Two Different Signal Plans. Transportation Research Record, 2674(7), 172-180. https://doi.org/10.1177/0361198120921161
Wang, Y., Wu, Z., Yang, X., & Huang, L. (2013). Design and implementation of an emergency vehicle signal preemption system based on cooperative vehicleinfrastructure technology. Advances in Mechanical Engineering, 5, 834976.
Wang, Y., Yao, Z., Cheng, Y., Jiang, Y., & Ran, B. (2021). Kalman filtering method for real-time queue length estimation in a connected vehicle environment. Transportation Research Record, 2675(10), 578-589.
Weinert, F., & Dring, M. (2015, 2015//). Development and Assessment of Cooperative V2X Applications for Emergency Vehicles in an Urban Environment Enabled by Behavioral Models. Modeling Mobility with Open Data, Cham.
Weinert, F., Dring, M., & Bogenberger, K. (2019, 27-30 Oct. 2019). Influence of Emergency Vehicle Preemption on Travelling Time and Traffic Safety in Urban Environments enabled by Innovative Behavioral Models and V2X Communication Simulation and Case Study. 2019 IEEE Intelligent Transportation Systems Conference (ITSC),
Wiedemann, R. (1974). Simulation des Strassenverkehrsflusses.
Wu, K., & Guler, S. I. (2019). Estimating the impacts of transit signal priority on intersection operations: A moving bottleneck approach. Transportation research part C: emerging technologies, 105, 346-358. https://doi.org/https://doi.org/10.1016/j.trc.2019.06.003
Xu, H., & Ying, J. (2017). Bus arrival time prediction with real-time and historic data. Cluster Computing, 20(4), 3099-3106. https://doi.org/10.1007/s10586-017-1006-1
Xu, T., Barman, S., Levin, M. W., Chen, R., & Li, T. (2022). Integrating public transit signal priority into max-pressure signal control: Methodology and simulation 349

study on a downtown network. Transportation research part C: emerging technologies, 138, 103614. https://doi.org/https://doi.org/10.1016/j.trc.2022.103614 Yang, K., Menendez, M., & Guler, S. I. (2019). Implementing transit signal priority in a connected vehicle environment with and without bus stops. Transportmetrica B: Transport Dynamics, 7(1), 423-445. https://doi.org/10.1080/21680566.2018.1434019 Yun, I., Park, B. B., Lee, C. K., & Oh, Y. T. (2012). Comparison of emergency vehicle preemption methods using a hardware-in-the-loop simulation. KSCE Journal of Civil Engineering, 16(6), 1057-1063. Zeng, X., Sun, X., Zhang, Y., & Quadrifoglio, L. (2015). Person-Based Adaptive Priority Signal Control with Connected-Vehicle Information. Transportation Research Record, 2487(1), 78-87. https://doi.org/10.3141/2487-07 Zhang, L., Gou, J., Liu, K., McHale, G., Ghaman, R., & Li, L. (2009, 20-23 Sept. 2009). Simulation Modeling and Application with Emergency Vehicle Presence in CORSIM. 2009 IEEE 70th Vehicular Technology Conference Fall, Zhao, J., Guo, Y., & Duan, X. (2017). Dynamic path planning of emergency vehicles based on travel time prediction. Journal of advanced transportation, 2017. Zhou, G., Gan, A., & Zhu, X. (2006). Determination of Optimal Detector Location for Transit Signal Priority with Queue Jumper Lanes. Transportation Research Record, 1978(1), 123-129. https://doi.org/10.1177/0361198106197800116
350