A Method for High-Value Driving Demonstration Data Generation Based on One-Dimensional Deep Convolutional Generative Adversarial Networks

Abstract

As a promising sequential decision-making algorithm, deep reinforcement learning (RL) has been applied in many fields. However, the related methods often demand a large amount of time before they can achieve acceptable performance. While learning from demonstration has greatly improved reinforcement learning efficiency, it poses some challenges. In the past, it has required collecting demonstration data from controllers (either human or controller). However, demonstration data are not always available in some sparse reward tasks. Most importantly, there exist unknown differences between agents and human experts in observing the environment. This means that not all of the human expert’s demonstration data conform to a Markov decision process (MDP). In this paper, a method of reinforcement learning from generated data (RLfGD) is presented, and consists of a generative model and a learning model. The generative model introduces a method to generate the demonstration data with a one-dimensional deep convolutional generative adversarial network. The learning model applies the demonstration data to the reinforcement learning process to greatly improve the effectiveness of training. Two complex traffic scenarios were tested to evaluate the proposed algorithm. The experimental results demonstrate that RLfGD is capable of obtaining higher scores more quickly than DDQN in both of two complex traffic scenarios. The performance of reinforcement learning algorithms can be greatly improved with this approach to sparse reward problems.