Aggregated Multi-deep Deterministic Policy Gradient for Self-driving Policy

Abstract

Self-driving is a significant application of deep reinforcement learning. We present a deep reinforcement learning algorithm for control policies of self-driving vehicles. This method aggregates multiple sub-policies based on the deep deterministic policy gradient algorithm and centralized experience replays. The aggregated policy converges to the optimal policy by aggregating those sub-policies. It helps reduce the training time largely since each sub-policy is trained with less time. Experimental results on the open racing car simulator platform demonstrates that the proposed algorithm is able to successfully learn control policies, with a good generalization performance. This method outperforms the deep deterministic policy gradient algorithm with 56.7% less training time.