Adaptive Transit Signal Priority based on Reinforcement Learning, Connected Vehicles and Software in the Loop Simulation
Author(s)
Kwesiga, Dickness Kakitahi
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This dissertation develops adaptive transit signal priority (TSP) algorithms based on reinforcement learning (RL), connected vehicles (CV), and software in the loop simulation (SILs). Before embarking on these rather advanced algorithms, sensitivity studies are carried out in a simulated environment to explore the fundamental principles of TSP including the selection of TSP strategies and evaluation of the critical factors and conditions that affect TSP performance. Starting from a single agent formulation for an isolated intersection, the study progresses to multi-agent formulations and tests RL-based adaptive TSP algorithms for real-world corridor implementation. The study develops multi-agent RL-based adaptive signal control before extending the algorithms to include TSP. The final corridor level adaptive signal control and adaptive TSP algorithms are based on multi-agent proximal policy optimization (MA-PPO) with centralized critic. The algorithms are developed in a traffic microscopic simulation environment. To address training feasibility, simulation and agent architecture decisions are undertaken to improve run-time efficiency. On a real-world based simulated seven-intersection test corridor, the formulated RL-based adaptive signal control algorithm is compared to the field optimal actuated signal control (ASC) by evaluating bus and mainline travel times and delay for side street movements. The field implemented ASC is modeled using SILs. Considering real time traffic and signal states, the RL-based TSP algorithms are formulated to adaptively select coordinated TSP phasing across multiple intersections. Two implementation architectures are considered for the trained TSP agents, (1) RL_RL_TSP system in which RL agents are always active to control background traffic and implement TSP when a bus approaches the intersection, and (2) RL_ASC_TSP system in which the background general traffic control runs ASC, and RL agents are event triggered to implement TSP when a bus approaches the intersection. On the test corridor, the two RL-based TSP systems show superior performance to the ASC check-in check-out (CI-CO) TSP system as evaluated by bus travel time. Both RL_RL_TSP and RL_ASC_TSP systems show marginal changes on the mainline travel time and side street delay for non-transit vehicles after TSP implementation while achieving delay reductions and schedule adherence improvements for the transit vehicle.
Sponsor
Date
2025-05-07
Extent
Resource Type
Text
Resource Subtype
Dissertation