Spillback-constrained cooperative ramp metering for incident-affected merge areas via constrained multi-agent reinforcement learning

Highway merge areas are vulnerable to incident-induced capacity drops, ramp queue-spillback, and congestion propagation. Existing deep reinforcement learning-based ramp metering methods mainly emphasize traffic efficiency and often treat queue limits as reward penalties, which may be insufficient under high-pressure incident conditions. This paper proposes a spillback-constrained cooperative ramp metering framework based on constrained multi-agent reinforcement learning. The control problem is formulated as a constrained Markov decision process, where ramp queue exceedance is modeled as an operational constraint. A Lagrangian-MAPPO algorithm balances traffic efficiency and spillback mitigation, while a graph attention network captures spatial interactions among ramps. A SUMO-based case study on the G2503 highway compares the proposed method with fixed-time control, ALINEA, and MADDPG. Results averaged over three independent random seeds show that the proposed framework improves ramp operations, suppresses queue accumulation, and enhances control stability under incident-induced disturbances.