A Reinforcement Learning Approach Based on Parameter Prediction for Multi-AUV Task Assignment in Ocean Current

In this paper, the reinforcement learning (RL) is used to study task assignment problem of multiple autonomous underwater vehicles (AUVs) in ocean current. To solve the problem that traditional RL is prone to trap into non-optimal policy in training, a RL algorithm based on parameter prediction is proposed. The algorithm consists of four parts: First, we design an objective function with four constraints, including ocean current, task emergency, energy consumption and collision detection. The AUVs need to balance these constraints to find optimal strategy. Second, the reward distribution matrix (RDM) is constructed by the team cumulative reward generated by multiple episodes. Third, the features in RDM are extracted by truncated singular value decomposition, and the features are used to predict the policy feature. Forth, the prediction feature is mapped into policy parameters by its surrounding features, and we design a recovery mechanism to avoid the prediction errors. Finally, we analyze the effectiveness of the algorithm, and the simulation results show that compared with DDQN, our algorithm have a better performance.