Discrete-time Finite Horizon Adaptive Dynamic Programming for Autonomous Vehicle Control

Nowadays, the autonomous vehicle control has become more and more critical, especially for the trajectory tracking problem. The existing model predictive control method applied to autonomous vehicle can reduce the tracking error. However, it requires online calculation for the model recurrence in long horizon, which causes heavy computational burden. This paper presents a novel discrete-time model predictive control solving framework. A finite horizon adaptive dynamic programming method was proposed to estimate the control policy for trajectory tracking. Under the constraint of self consistent condition, the control policy is fitted through deep neural network as a parameterized function to obtain faster solution. The performance of the proposed method is tested and verified under double lane change conditions via CarSim. The experimental results show that the system can accurately track given reference trajectory with 5000Hz frequency, which is three times faster than that solved by least squares estimation method.