An Autonomous Driving Decision Making Method with Maximum Entropy Objective and Safety Constraints

At present, the algorithm based on reinforcement learning has shown advantages over the rule-based algorithm in the field of building autonomous driving decision-making model, which has great potential and application prospects in the field of autonomous driving. However, reinforcement learning based algorithms often have disadvantages such as low learning efficiency, poor safety, and sensitivity to hyperparameters. To solve this problem, based on the Soft Actor-Critic algorithm with maximum entropy objective, the safety rule constraint is introduced to adjust the action sampling probability in the training phase, and the autonomous driving decision-making model under the conditions of 3 actions and 5 actions is trained and tested. It was compared with the value-based ssafety constraint algorithm and the rule-based safety constraint algorithm using DQN algorithm. The results show that the proposed algorithm is superior to the compared algorithms in terms of safety and training efficiency, and the algorithm based on SAC and safety constraints can improve the safety and learning efficiency of reinforcement learning algorithms.