Application of DDPG Algorithm for Residual Structured Networks in Autonomous Driving

The development of intelligent mobility and the spread of artificial intelligence have contributed to the widespread use of autonomous driving techniques. Among the various algorithms used in this field, Deep Deterministic Policy Gradient (DDPG) algorithm is favoured by researchers. The DDPG algorithm may ignore some features in the deep learning of the network, slow updating of parameters, and lack of stability and learning efficiency of the algorithm. To overcome these problems, this study proposes the deep deterministic policy gradient (R-DDPG) algorithm with the introduction of residual blocks. By establishing an experience pool of successful samples, the algorithm improves the sampling utilisation, enhances the exploration strategy and improves the neural network training efficiency. The performance of the R-DDPG algorithm is evaluated on the TORCS racing simulation platform, which demonstrates that the algorithm is able to improve the learning efficiency of driving strategies in an autonomous driving task compared to existing algorithms.