Deep Reinforcement Learning for Tuning of Adaptive Model Predictive Control for Autonomous Driving

Model Predictive Control (MPC) has emerged as a pivotal technology for optimizing control tasks in autonomous driving, particularly within Adaptive Cruise Control (ACC) systems. However, the manual tuning of MPC cost function weights and prediction horizons remains a significant challenge. In this paper, we introduce a novel framework that combines Deep Reinforcement Learning (DRL) with MFC to dynamically tune both the weight parameters and prediction horizon in real time. This approach, referred to as the Weights and Prediction Horizon Varying MPC (W-PH-MPC), overcomes traditional MPC limitations by utilizing proximal Policy optimisation and Deep Deterministic Policy Gradient (DDPG) algorithms to adjust control parameters. We evaluate the effectiveness of our approach through simulations in vehicle-tracking scenarios. Simulation results show that the adaptive MPC-RL controller achieves better tracking performance, without compromising power consumption, and lowers longitudinal jerk compared to a fixed-parameter MPC baseline, resulting in smoother and more efficient vehicle behavior.