An AUV Target-Tracking Method Combining Imitation Learning and Deep Reinforcement Learning

This study aims to solve the problem of sparse reward and local convergence when using a reinforcement learning algorithm as controller an AUV. Based on generative adversarial imitation (GAIL) combined with multi-agent, multi-agent GAIL (MAG) is proposed. The enables AUV directly learn from expert demonstrations, overcoming difficulty slow initial training network. Parallel multi-agents reduces high correlation between samples avoid convergence. In addition, function designed help training. Finally, results show that in unity simulation platform test, proposed has strong optimal decision-making ability tracking process.