Resilient Multi-Agent Deep Reinforcement Learning for Routing Control via Traffic Cost Fields

Urban road networks frequently experience recurrent congestion and non-recurrent disruptions, where individually optimal routing can degrade system-level efficiency under mixed compliance. Naive shortest-path guidance can be overly myopic in coupled traffic systems. As congestion is shaped by collective route choices, it may over-react to momentary link costs and trigger herding-like oscillations. This motivates routing controllers with decentralized execution that scale to many vehicles while better accounting for congestion externalities. We propose a routing framework centered on Traffic Cost Fields (TCF), which maintain factorized edge-wise signals as an interpretable interface between traffic states and routing decisions. A shared-parameter multi-agent actor-critic policy is trained with Proximal Policy Optimization (PPO) to make periodic intersection-level next-hop decisions from masked local observations, enabling decentralized execution at scale. Experiments on an abstracted road graph extracted from Shenzhen show improved average and tail travel-time performance over shortest-path baselines across multiple demand levels.