A multi-agent reinforcement learning framework for integrated traffic signal control and dynamic bus lane access management

• An MARL-based system jointly controls signals and vehicle access to bus lanes. • Agents co-evolve through centralized training, decentralized execution. • TransitStream outperforms SotA methods under both peak and off-peak scenarios. • Co-evolution improves performance against independently trained control components. • Mixed traffic in bus lanes may enhance transit performance by preventing gridlocks. This paper presents TransitStream, a novel multimodal traffic management strategy based on Multi-Agent Reinforcement Learning (MARL), designed to enhance transit operations on urban arterials, while maintaining optimal service levels for private vehicles. TransitStream employs a dual-control mechanism, integrating traffic signal control and transit priority lane density control via Variable Message Signs (VMS). The density control mechanism aims at regulating the access of private vehicles into transit priority lanes. Traffic signal control on the other hand, focuses on reducing the waiting times of vehicles at intersections, completely disregarding the effect on public transport. While seemingly competitive, these objectives’ interdependence leads to a coevolution of mutually beneficial policies during the training process of the MARL framework by utilizing the popular Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. TransitStream is tested in SUMO micro-simulation software, on a complex arterial with high traffic and transit demands in the city of Athens (Greece). Findings reveal that a certain level of mixed traffic can be beneficial for the overall public transport operations, even though it can deteriorate the performance of some individual buses, as it prevents gridlock situations and maintains smoother traffic flow. An extensive comparative analysis against state-of-the-art traffic management strategies demonstrated that TransitStream consistently outperformed existing transit prioritization methods based on space segregation between modes and traffic signal control, as well as combinations of the above strategies.