Autonomous Driving Decision-Making Based on an Improved Actor-Critic Algorithm

Autonomous driving technology, as a new type of automotive driving technology, contributes to reducing the number of traffic accidents and to lowering the mortality rate for traffic accidents. However, due to the limited prior knowledge of designers, the current autonomous driving decision-making systems face difficulties in dealing with complex and everchanging traffic scenarios. In view of this, this study proposes an autonomous driving decision-making model based on the Soft Actor-Critic algorithm and a long short-term memory network. The experimental results obtained in complex mixed traffic scenarios for the average collision frequency, the average lane change frequency, the average following distance, and the average distance from the lane centerline for the decision-making model based on the Soft Actor-Critic algorithm and a long short-term memory network were 0.9, 9.8, 19 metres, and 0.36 metres, respectively. Moreover, the average arrival time, root mean square of acceleration, and root mean square of the acceleration change rate obtained by the proposed model were 127 seconds, 1.1, and 1.3, respectively, these values being superior to the ones obtained by the other employed models. In addition, when facing a sudden pedestrian crossing, the collision time for the proposed decision-making model was the shortest at 8.3 seconds, which is 1.2 seconds higher than the values obtained by the other employed models. The obtained outcomes prove that the decision-making model based on the Soft Actor-Critic algorithm and a long short-term memory network proposed in this paper can cope with complex and changing traffic scenarios, and ensure the safety of both pedestrians and drivers.