Reinforcement Learning–Based Ramp Metering Strategy Considering Queue Management

Abstract

This paper introduces an action replacement module for reinforcement learning (RL)–based ramp metering to address the issue of ramp queue spillback during the training process. Ramp queue spillback leads to significant impacts on the traffic efficiency of adjacent road networks, making it a critical concern in ramp control. Existing RL approaches often employ ramp states as reward functions to encourage agents to learn strategies that avoid queue overflow. However, due to the trial-and-error nature of RL, these methods frequently generate actions that cause queue spillback during training, posing challenges for real-time online training in real-world applications. To overcome this limitation, the proposed action replacement module utilizes the store-and-forward model to estimate a lower bound for ramp metering rates. By identifying and replacing actions that fail to meet this constraint, the strategy effectively prevents queue spillback. In addition, penalties are imposed on replaced actions to guide the agent in learning effective and practical control policies. The proposed method is evaluated in both single-ramp and multiramp scenarios. Experimental results demonstrate that the agent can learn the queue spillback prevention strategies, and nearly eliminate ramp queue spillback without compromising control performance.