SAC-Based Collision Avoidance Control for Vehicles by the Efficient Experience Replay

With the application of learning-based methods, au-tonomous driving technology is becoming increasingly powerful. Reinforcement learning (RL) is being widely adopted in the field for driving safety. However, a major challenge in using RL methods is their difficulty in convergence and long training cycles. This paper proposes an improved SAC (Soft Actor-Critic) algorithm, based on an efficient experience replay mechanism. The experience replay mechanism combines directional sampling and uniform random sampling during experience replay. We applied the proposed improved SAC algorithm to train intelligent vehicles in lead vehicle braking scenarios and lead vehicle lane changing scenarios and evaluated its performance. The results indicate that the proposed improved SAC algorithm accelerates the convergence rate by an average of 23.1 % compared to the SAC algorithm, and increases the average reward value after stable convergence by 1.16 %. In safety tests, the collision occurrence rate was 0.1 %.