Fast Solution of Adaptive Dynamic Programming for Intelligent Vehicle Predictive Controller

In this paper, we propose an offline Deep Adaptive Dynamic Programming (DADP) algorithm to improve the real-time performance of trajectory tracking control. Model Predictive Control (MPC) is an effective method for solving nonlinear optimal control problems of trajectory tracking, while as an online solution, the processing speed of MPC is too slow to meet the high real-time requirements of automatic driving motion control, especially under long predictive step length. DADP consists of two parts, policy evaluation and policy improvement, where both the policy function and the evaluation function are approximated by neural networks mapping the state of agent to control outputs and value evaluation of the state respectively. Besides, multi-agent parallel training is adopted to speed up the neural network training. The optimal policy function is obtained after the iterations of policy evaluation and policy improvement reach the final equilibrium. Finally, double lane change tracking is performed on the CarSim digital twin platform verifying the high real-time performance of DADP.