Scalable and Reliable Multi-Agent Reinforcement Learning for Traffic Assignment

Escalating urbanization and increased travel demand impose stringent benchmarks on traffic assignment methodologies. Multi-agent reinforcement learning (MARL) approaches outperform traditional methods in emulating adaptive travel routing without requiring explicit system dynamics, which is beneficial for real-world implementation. Nevertheless, MARL frameworks face challenges in scalability and reliability when managing extensive networks with substantial travel demand, which restricts their practical use in solving large-scale traffic assignment problems. This research introduces an innovative MARL framework for traffic assignment, redefining agents as origin-destination (OD) routers instead of individual travelers, enhancing scalability. Additionally, a specialized action space formulation using a proposed Dirichlet-based strategy and a reward formulation based on the local relative gap is crafted to efficiently reach optimal solutions, increasing model reliability. Experiments demonstrate the proposed MARL framework effectively handles medium-sized networks with extensive and varied city-level OD demand, surpassing existing MARL methods. Applied to the SiouxFalls network, the method achieves better assignment outcomes in fewer steps, reducing the relative gap by 83.6 % compared to traditional techniques.