A Deep Reinforcement Learning Method for Self-driving

Self-driving technology is an important issue of artificial intelligence. Basing on the end-to-end architecture, deep reinforcement learning has been applied to research for self-driving. However, self-driving environment yields sparse rewards when using deep reinforcement learning, resulting in local optimum to network training. As a result, the self-driving vehicle does not obtain correct actions from outputs of neural network. This paper proposes a deep reinforcement learning method for self-driving. According to the classification threshold value that is dynamically adjusted by reward distributions, the sparse rewards is divided into three groups. The experience information for different rewards is fully utilized and the local optimum problem in the network training process is avoided. By comparing with the traditional method, simulation results show that the proposed method significantly reduces the training time of network.