Development of Applicable Reinforcement Learning Compensator Using Ranking Information for AUV

In nonlinear multi input multi output systems such as Autonomous Underwater Vehicle (AUV), a physical dynamical model is usually used to design a control system. To satisfy trade-off between plural control targets, design and adjustment of the control system are complicated, and which lead to struggle by control design expert. In recent years, applying data-driven techniques such as deep reinforcement learning have been tried to improve control performance by automatic adjustment in the deep neural network. However, we have difficulty to applying data-driven methods in sight of designing reward functions, and how to combine learning results into a controller so as not to lose prior information such as models. Then, in this study, we developed compensation control technology by deep reinforcement learning (DRL) for efficient control under multi-purpose trade-off, by using selector with ranking to several options of control target trajectory, and to verify its effectiveness by simulation using obstacle avoidance as an example.