MARL-based cooperative transit signal priority for the arterial road to reduce schedule delay

Abstract

Transit signal priority (TSP) is an effective strategy to reduce transit delays and improve intersection efficiency. This paper introduces a Cooperative TSP strategy of Variable phase (CTSPV) using multi-agent reinforcement learning (MARL) to minimize transit schedule delays on arterial roads. The agents adjust phase sequences and durations based on real-time traffic, balancing transit and non-transit vehicle needs, resolving conflicting bus requests, and ensuring agent cooperation. Invalid action masking ensures compliance with green time and phase-skipping rules. Simulation results show CTSPV reduces person delay, queue lengths, and lateness by 8.7%, 31.6%, and 17.0%, respectively, compared to fixed-time signals. Testing different green time constraints highlights the importance of proper restrictions for efficient learning. Analysis of CTSPV's signal timing reveals agents prioritize phases with high traffic demand and bus priority, skipping phases with lower demand. Evaluation results of generalized rule-based strategies based on those RL-derived patterns demonstrate the good performance of RL-learned knowledge.