Computational modeling demonstrates robust vehicle platoon control during communication and actuator failures, highlighting improved stability in mixed traffic.
Key Points
Develop a knowledge-guided deep reinforcement learning framework that maintains stable and adaptive connected vehicle platoon coordination during communication delays and actuator failures.
Designed a hierarchical framework combining proximal policy optimization (PPO), a robust consensus controller, and a control-error-sensitive graph attention network (CEGAT).
Embedded control stability conditions directly into training via policy guidance, continuous reward shaping, and linear matrix inequality (LMI) screening before deployment.
Evaluated the model using naturalistic driving data from the OpenACC database across heterogeneous platoons up to 16 vehicles with varied human-driven vehicle placements.
Knowledge guidance cut training time by 50% relative to standard reinforcement learning without domain guidance.
Reduced tracking errors by up to 32% compared to baseline methods and improved velocity tracking accuracy by over 57% relative to commercial adaptive cruise control.
Maintained robust platoon stability with human-driven vehicle proportions up to 62.5% and scaled seamlessly up to 16-vehicle formations regardless of vehicle placement.