Highway Merging Control Using Multi-Agent Reinforcement Learning

This paper presents a multi-agent reinforcement learning approach for autonomous vehicle highway merging control. A decentralized partially observable Markov decision process is formulated, where each autonomous vehicle acts independently based on local observations. The scenario considered in this paper assumes randomly spawning vehicles and fluctuating traffic flows and a self-attention network is used to handle varying numbers of agents (vehicles). The proposed method is validated in SUMO traffic simulator, which provides a realistic highway simulation environment. Results demonstrate the approach can enable safe, efficient coordination for merging maneuvers, successfully handling dynamic number of agents. Future work will continue to enhance multi-agent reinforcement learning for autonomous vehicle coordination in complex traffic environments by reducing the training time.