关键词: 车辆安全, 自主避撞, 深度确定性策略梯度(DDPG), 控制模型, 多目标奖励函数

Abstract:

A vehicle autonomous collision avoidance decision control model was proposed based on a deep deterministic policy gradient (DDPG) to improve the self-learning and decision-making capabilities of vehicle in driving environment. A state space containing self-vehicle and target object information, and an action space including the self-vehicle deceleration were designed based on a reinforcement learning theory of Markov decision process and a longitudinal kinematic of vehicle. An end-to-end vehicle autonomous collision avoidance decision model was constructed which takes safety, comfort and efficiency into a multi-objective reward function. An interaction model was built by using MATLAB/Simulink with the DDPG algorithm and the traffic environment, and the model passed through test for scenarios of car to car stationary (CCRs) and scenarios of car to car braking (CCRb). The results show that the proposed decision-making algorithm has good convergence with introducing limit values of acceleration and jerk, realizes the effective collision avoidance of vehicle with considering ride comfort. Therefore, it has better performances than using fuzzy control.

Key words: vehicle safety, autonomous collision avoidance, deep deterministic policy gradient (DDPG), control model, multi-objective reward function