A dynamic self-improving ramp metering algorithm based on multi-agent deep reinforcement learning

We present a novel ramp metering algorithm that incorporates multi-agent deep reinforcement learning (DRL) techniques, which utilizes monitoring data from loop detectors. Our proposed approach employed a multi-agent DRL framework to generate optimized ramp metering schedules for each ramp meter in real-time, enhancing the operational efficiency of urban freeways with less investment. To simplify the implementation and training of the algorithm, we developed a simulation platform based on SUMO microscopic traffic simulator. We conducted a series of simulation experiments, including local and coordinated ramp metering scenarios with various traffic demands profiles. The simulation results indicate that the proposed DRL-based algorithm outperforms the state-of-the-practice ramp metering methods, considering a comprehensive evaluation index encompassing mainstream speed at the bottleneck and queue length on ramp. Additionally, the method exhibits robustness, scalability, and the potential for further improvement through online learning during implementation.