Integrating StateoftheArt Mobilityon Demand Fleet Models into Transportation

Integrating State-of-the-Art Mobility-on. Demand Fleet Models into Transportation System Simulation Tools for Policy Analysis MIC HAE L HYL AND UNIVE RSIT Y OF CALIF OR NIA IRVINE HYL AN DM@UCI. EDU Workshop III: Large Scale Autonomy: Connectivity and Mobility Networks

Acknowledgements PROJECT 1 PROJECT 2 Collaborators ◦ UCI: Navjyoth Sarma, Daisik Nam, Dingtong Yang, Arash Ghaffar ◦ Argonne National Laboratory: Felipe De Souza, I. Omer Verbas ◦ UCI: Steve Ritchie, Craig Rindt, R. Jayakrishnan, Mike Mc. Nally, Wenlong Jin ◦ San Diego Association of Governments Funding ◦ Argonne National Laboratory ◦ NSF Smart and Connected Communities HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 2

Outline/Agenda Background ◦ Transportation System Simulation Modeling Software ◦ (Automated) Mobility-on-Demand Fleet Modeling Context – Research Project 2 ◦ Plans to create the next generation of regional transportation planning models Focus – Research Project 1 ◦ ◦ More on Mobility-on-Demand Fleet Models The Key Trade-off: Fleet Performance vs. Computational Efficiency Solution strategies that are operationally effective and computationally efficient Results HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 3

Transportation System Simulation Models Main Characteristics ◦ Agent-based with Synthetic Populations ◦ ◦ Socio-demographic Information Vehicle Ownership Home + Work Locations Activities to Complete ◦ Activity-based Travel Demand Models ◦ Activity Location and Time Choice ◦ Mode Choice for Tours and Trips ◦ Dynamic Network Assignment-Simulation Models ◦ Route Choice (and Departure Time Choice) in Network ◦ Congestion is Endogenous CT RAMP NU-TRANS HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 4

Transportation System Simulation Models Why? ◦ Model the complexities of travel behavior and transportation network congestion ◦ Provide useful insights to transportation decision-makers ◦ Forecast future scenarios ◦ Analyze Infrastructure Investments ◦ Analyze Transportation + Land-use Regulations CT RAMP NU-TRANS HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 5

Transportation System Simulation Models Benefits (compared to analytical models) ◦ High-fidelity ◦ Extensions and Integration ’can be straightforward’ ◦ At least computationally ◦ If not, behaviorally ◦ Policy-sensitive ◦ Large Volume of Output Data Disadvantages ◦ Data-intensive ◦ Calibration and Validation can be very difficult CT RAMP ◦ Theoretical Guarantees mostly non-existent NU-TRANS HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 6

Transportation System Simulation Models A New Addition to Transportation System Simulation Models! Mobility Services (e. g. Uber and Lyft) ◦ ◦ ◦ Mobility-on-Demand Automated Mobility-on-Demand Dynamic Ride-sharing Micro-transit Etc. Vehicle Fleet Models – Dynamic Dial-a-ride-Problem ◦ ◦ Dynamic Traveler-Vehicle Matching/Assignment Dynamic Routing Dynamic Scheduling Dynamic Repositioning CT RAMP NU-TRANS HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 7

Broader Research Study ◦ Addressing Unprecedented Community-Centered Transportation Infrastructure Needs and Policies for the Mobility Revolution ◦ National Science Foundation (NSF): Smart and Connected Communities Planning Grant HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 8

Project Goals Improve communities: ◦ Sustainability ◦ Livability/Accessibility ◦ Mobility Through decision support: ◦ Infrastructure Upgrades ◦ Transport Policies/Regulations Related to new ‘innovations’ ◦ Mobility Service Providers (MSPs) ◦ Connected Automated Vehicles (CAV) This Photo by Unknown Author is licensed under CC BY-ND HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 9

Problem Statement SOCIETAL PROBLEMS TECHNOLOGICAL CHALLENGES Viability of MSPs and CAVs is uncertain Incorporating MSPs into transport planning models ◦ Financially ◦ Technologically ◦ Politically Need for Planning and Regulating around MSPs and CAVs to… ◦ Obtain their societal benefits ◦ Avoid pitfalls Optimizing transport policies and infrastructure decisions ◦ Related to CAVs and MSPs ◦ For community outcomes! HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 10

Old Transportation System (Model) Infrastructure Supply Roads, Lanes, Intersections, Tolls, Communications, Transit Lines, Bus Stops, etc. Transport Infrastructure Managers Travel Demand Route Choices, Mode Choices, Departure Time Choices, Activity Choices, Location Choices, etc. HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 11

New Transportation System (Model) Infrastructure Supply Roads, Lanes, Intersections, Tolls, Communications, Transit Lines, Bus Stops, etc. Transport Infrastructure Managers MSPs with (CAV) Fleets Service Design Sharing, Fleet Size, Prices, Door-to-Door vs. Pick/Drop Points Vehicle Design Vehicle Size, Seat Arrangement, Sensors, Controllers Travelers HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS Travel Demand Route Choices, Mode Choices, Departure Time Choices, Activity Choices, Location Choices, etc. 12

Optimization-based Transport Decisions HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 13

Research Questions How are MSPs currently impacting transport systems? What are potential impacts of specific policies on transport systems with MSPs and CAVs? What are socially-optimal CAV-related infrastructure upgrades ◦ What is the optimal deployment timeline for the infrastructure upgrades? This Photo by Unknown Author is licensed under CC BY-SA HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 14

Integrating Dynamic Mobility-on-demand Fleet Models into Transportation System Simulation Models A BALANCING ACT BETWEEN COMPUTATIONAL EFFICIENCY AND FLEET PERFORMANCE HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 15

Introduction Automated mobility-on-demand (AMOD) services ◦ Fleet of driverless vehicles ◦ Serving passenger transport requests, on-demand ◦ Fleet controlled by a single entity ◦ Probably centrally controlled ◦ Probably owned by the entity as well ◦ Initially, entities in the US appear to be private companies This Photo by Unknown Author is licensed under CC BY-SA Examples ◦ Waymo in Arizona ◦ ~10 fully driverless robo-taxis ◦ Hyundai/Pony. AI in Irvine, CA ◦ Still includes safety backup driver https: //www. ocweekly. com/riding-in-self-driving-pony-ai-cars-through-irvine/ HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 16

AMOD Service Definition ◦ ◦ ◦ Travelers request rides dynamically via a mobile application Travelers want to be served (i. e. picked up) immediately AVs transport travelers directly from their requested pickup location to their drop-off location AVs in the fleet are functionally homogeneous The fleet size is fixed in the short term (i. e. one-day) The fleet operator has complete control over each AV HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 17

Background 1. Significant research on Optimal Dispatching for MOD/AMOD vehicle fleets ◦ Optimization-based Assignment Policies: ◦ Hyland Mahmassani (2018), Hörl et al. (2019), Dandl et al. (2019), Bertsimas, Jaillet, and Martin (2019) ◦ (Multi-Agent) Reinforcement Learning: ◦ Oda and Tachibana (2018), Holler et al. (2019), Li et al. (2019), Singh, Alabbasi, and Aggarwal (2019) ◦ Approximate Dynamic Programming: ◦ Al-Kanj, Nascimento, and Powell (2018) HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 18

Background 2. Significant research analyzing the Transportation System Impacts of MOD/AMOD vehicle fleets ◦ Rule-based AMOD Fleet Control Policies implemented in Agent-based Transportation System Simulation Software ◦ Nearest Neighbor: Zhang et al. (2015) ◦ Smart Nearest Neighbor: Bischoff and Maciejewski (2016) ◦ Zone Partitioning: Fagnant and Kockelman (2014, 2015, 2016), Boesch et al. (2016), Gurumurthy et al. (2020) HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 19

Research Problem and Objective Research Problem ◦ Need to model real-world AMOD fleet with high-quality dispatching policy within agent-based transportation simulation model ◦ Current AMOD fleet dispatching policy dichotomy: 1. 2. Computationally efficient but operationally ineffective policies Computationally inefficient (for agent-based simulation models) but operationally effective policies Research Objective ◦ Develop a AMOD fleet dispatching policy that is: ◦ Computationally efficient enough to handle large-scale AMOD fleets (e. g. 2, 000 – 20, 000 vehicles) in a dynamic network model ◦ Operationally effective enough to mimic a real-world AMOD fleet HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 20

AMOD Operational Problem Components 1. Assignment/Matching of requests to AVs 2. Routing and Scheduling vehicles to pickup and drop-off requests 3. Repositioning AVs to balance supply and future demand Comments ◦ Routing and Scheduling are not difficult because there is no ride-sharing, and requests want service immediately ◦ Assignment-Routing-Repositioning can be addressed simultaneously or sequentially This study focuses on Assignment only HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 21

Dynamic Assignment Problem Statement HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 22

Problem Formulation: Objective HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 23

Problem Formulation: Constraints HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 24

Reducing Computation Complexity: #1 HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 25

Reducing Computation Complexity: #2 HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 26

Reducing Computation Complexity: #3 HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 27

Reducing Computation Complexity: #4 This Photo by Unknown Author is licensed under CC BY-SA HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS This Photo by Unknown Author is licensed under CC BY-SA 28

Reducing Computational Complexity Example Heap (vid, distance) -inf z 1 z 2 z 3 z 7 z 8 z 4 8 z 6 1 z 12 z 16 z 17 z 21 z 22 2 z 13 5 z 5 7 z 9 z 10 TNC vehicles z 14 3 z 15 4 Eligible vehicles z 18 z 19 z 20 z 23 z 24 z 25 HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 29

Reducing Computational Complexity Example Heap (vid, distance) (1, 5) Heap graph for a passenger (1, 5 ) (2, 7) z 1 z 2 z 3 z 7 z 8 z 4 8 z 6 1 z 12 z 16 z 17 z 21 z 22 2 z 13 5 z 5 7 z 9 z 10 TNC vehicles z 14 3 z 15 4 Eligible vehicles z 18 z 19 z 20 z 23 z 24 z 25 HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 30

Reducing Computational Complexity Example Heap (vid, distance) (1, 5) Heap graph for a passenger (1, 5 ) (2, 7) z 1 (3, 6) z 2 z 3 z 7 z 8 z 4 8 z 6 1 (3, 6 ) (4, 8) z 11 z 12 z 16 z 17 z 21 z 22 2 z 13 5 z 5 7 z 9 z 10 TNC vehicles z 14 3 z 15 4 Eligible vehicles z 18 z 19 z 20 z 23 z 24 z 25 (4, 8) HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 31

Reducing Computational Complexity Example Heap (vid, distance) (5, 4) Heap graph for a passenger (5, 4 ) (2, 7 ) z 1 (3, 6) z 2 z 3 z 7 z 8 (2, 7) z 4 8 z 6 1 (3, 6 ) (1, 5) z 11 z 12 z 16 z 17 z 21 z 22 2 z 13 5 (4, 8) z 5 7 z 9 z 10 TNC vehicles z 14 3 z 15 4 Eligible vehicles z 18 z 19 z 20 z 23 z 24 z 25 (4, 8) HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 32

Implementation C++ Dynamically Linked Library (DLL) connects ◦ AMOD Dispatching Module to ◦ POLARIS Agent Based Transportation Simulation Software ◦ Developed by Argonne National Labs Optimization Package: CPLEX 12. 8 Parameters: ◦ Max Pickup time (M) = 20 min ◦ Nearest k- vehicles/requests parameter k = 10 ◦ Switch-point from Point-to-Point to Zone-to-Zone costs = 15% of M ◦ Base Fleet Size = 1000 vehicles Network: Bloomington, IL ◦ 2, 540 Nodes ◦ 7, 023 Links ◦ 185 Zones ◦ Simulation Period: 24 hours ◦ Fleet size: ◦ ◦ Base Scenario: 1, 0000 vehicles Other scenarios: 500 to 3, 000 vehicles Figure 1: Bloomington, IL Network HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 33

Dispatching Policy Comparison and Evaluation Measures DISPATCHING POLICIES EVALUATION MEASURES # Policy Optimizationbased? K-Nearest? Cost Objects A 1 Rule-based Zonal Policy No NA Zonal skims Optimal, k-Nearest Hybrid, Zonal Skims Yes A 2 Optimal, k-Nearest Hybrid, Point-to-Point A 3 and Zonal Skim Hybrid A 4 Unrestricted Assign. Problem with Pointto-Point Shortest Path Total Requests Assigned across 24 hours Average Pickup distance (feet) Fleet Productivity (%): Yes Zonal Skims Yes Point-to. Point SP and Zonal Skims Yes No ◦ The proportion of fleet miles in which the vehicles carry a passenger POLARIS Run Time (seconds) Point-to. Point SP HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 34

Results: Base Scenario Policy A 1 A 2 A 3 A 4 Fleet Size 1000 Assigned Travelers 92, 162 101, 504 103, 464 105, 934 Polaris Run Time 1104 (sec) 1138 1291 92, 960 Avg Pickup Distance (feet) 1, 994 1, 855 1, 271 76. 7 78. 5 83. 3 3, 021 Fleet 68. 4 Productivity (%) A 3 significantly improves efficiency compared to global optimal policy, A 4: ◦ 98% reduction in POLARIS run time ◦ 5%-point reduction in fleet productivity ◦ 50% increase in average pickup distance A 3 significantly improves effectiveness compared to rule-based policy, A 1: ◦ 12% increase in assigned travelers ◦ 40% reduction in average pickup distance A 3 improves effectiveness compared to using Zonal cost objects only policy, A 2: ◦ 7% reduction in average pickup distance HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 35

Results: Average Pickup Distance HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 36

Results: Polaris Run Time HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 37

Conclusion Agent-based Transportation System Simulation Models have many benefits, including: ◦ Integration of new modeling components, such as AMOD service fleet models ◦ That allows for policy analysis related to AMOD service models Model Fidelity is Important for Policy Analysis, but Computation Time also Matters ◦ This study proposes a method to significantly decrease computation time (~98%), while still reasonably capturing efficiency of MOD service Future Work ◦ Optimize Policy Decisions and Infrastructure Investments using: ◦ Transportation System Simulation Model for ‘Evaluation’ ◦ Simulation-Optimization Techniques ◦ Multi-Level Math Programming HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 38

Thank You! MICHAEL HYLAND -- HYLANDM@UCI. EDU

Introduction Benefits of Driverless Vehicles ◦ Economic ◦ Eliminate driver costs significantly reduce operational costs ◦ Operational ◦ Centralized (Full) control over entire vehicle fleet significant operational efficiency benefits This Photo by Unknown Author is licensed under CC BY-ND HYLAND - IPAM 2020 WSIII: LARGE SCALE AUTONOMY: CONNECTIVITY AND MOBILITY NETWORKS 40