Learning-Based Adaptive Optimal Control for Connected Vehicles in Mixed Traffic: Robustness to Driver Reaction Time

Area of Science:

• Intelligent Transportation Systems
• Control Theory
• Artificial Intelligence

Background:

• Connected and autonomous vehicles (CAVs) offer enhanced traffic safety and efficiency through vehicle-to-vehicle (V2V) communication.
• Heterogeneous driver behaviors and human-vehicle interactions pose significant challenges for designing effective control strategies for CAVs in mixed traffic platoons.

Purpose of the Study:

• To propose an adaptive optimal control design for a CAV at the tail of a platoon with multiple preceding human-driven vehicles.
• To enable the CAV controller to adapt to unknown and heterogeneous driver-dependent parameters without prior knowledge of individual driver car-following models.

Main Methods:

• Utilizing reinforcement learning and adaptive dynamic programming techniques for near-optimal controller learning from real-time V2V data.
• Employing an off-policy learning algorithm that leverages both historical and real-time data to enhance safety and reduce learning duration.

Main Results:

• Demonstrated the ability of the CAV controller to adapt to varying platoon dynamics caused by diverse driver behaviors.
• Achieved a considerable reduction in learning time through the proposed off-policy learning strategy.
• Validated the effectiveness and efficiency of the adaptive control method via rigorous theoretical proofs and microscopic traffic simulations.

Conclusions:

• The proposed adaptive optimal control method effectively manages CAVs in mixed-traffic platoons by learning from real-time V2V data.
• The approach enhances traffic safety and efficiency by adapting to unknown human driver behaviors, paving the way for more robust autonomous driving systems.