Optimal control of CAVs in mixed traffic environments
To address the challenges posed by the stochastic nature of human-driven behaviors, the heterogeneity in traffic composition, and the partial observability characteristic of complex mixed traffic conditions involving both CAVs and HVs, optimizing CAV driving behavior from a systemic traffic flow perspective remains particularly difficult. In response to these challenges, we have conducted extensive research on decision-making control for CAVs in mixed traffic environments. By integrating the core principles of physics models, control theory, and traffic flow theory into a deep reinforcement learning (DRL) framework, I developed the physics-informed deep reinforcement learning (PIDRL) methodology, specifically designed for the intelligent control of CAVs and other ITS agents in mixed traffic scenarios. This generic methodology has proven versatile and effective in various applications, including CAV control and intelligent bus operations. It has been validated in improving key system-level performance metrics such as safety, efficiency, stability, and energy utilization.