GACS: Generative Adversarial Imitation Learning Based on Control Sharing

Generative adversarial imitation learning (GAIL) directly imitates the behavior of experts from human demonstration instead designing explicit reward signals like reinforcement learning. Meanwhile, GAIL overcomes defects traditional by using a generative adversary network framework and shows excellent performance in many fields. However, acts on immediate rewards, feature that is reflected value function after period accumulation. Thus, when faced with complex practical problems, efficiency often extremely low policy may be slow to learn. One way solve this problem guide action (policy) agents' process, such as control sharing (CS) method. This paper combines proposes novel called based (GACS). GACS learns model constraints expert samples uses networks directly. The actions are produced used optimize effectively improve efficiency. Experiments autonomous driving environment real-time strategy game breakout show has better generalization capabilities, more efficient experts, can learn policies relative other frameworks.