Learning-Based Model Predictive Control for Vehicles with Modeling Bias

The dynamic uncertainty system has modeling bias, which has led to an increasing interest in learning-based model-driven control technology. In order to integrate the advantages of self-exploration of reinforcement learning (RL) and high robustness of model predictive control (MPC), we propose a learning-based model predictive control (LMPC) method. This algorithm is a method that can find the optimal control strategy under the guidance of simplified model and fuzzy cost function. The results show that taking vehicle motion control as an example. By adjusting the variable horizon, the vehicle model built can quickly and effectively achieve the control objectives. Compared with DDQN, LMPC reduces the disorder enumeration of RL, accelerates the optimization process of the control system to the target, and makes the exploration process more stable. Compared with MPC, it solves the problem that the control accuracy of the control system is not high and the model and parameters cannot be adjusted adaptively under the simplified modeling.