Deep Reinforcement Learning-Based Multi-AUV Task Allocation Algorithm in Underwater Wireless Sensor Networks

Autonomous underwater vehicle (AUV)-based data collection can bring significant advantages to underwater wireless sensor networks (UWSNs). Collaborative collection based on multi-AUV task allocation is an effective way to reduce delay. However, the existing research work seldom considers the real current environment in the task allocation, which leads to the large delay and yaw of AUVs. The introduction of ocean currents makes the existing task allocation algorithms no longer applicable due to the poor solving ability and long convergence time. Therefore, we propose an efficient task allocation algorithm named genetic algorithm N-step reinforcement learning improved DQN (GA-NDQN) by combining the genetic algorithm (GA) and N-step reinforcement learning (RL) nature of DQN to minimize the data collection delay. In our work, to minimize the impact of ocean currents on the AUV’s travel, the specific trajectory optimization problem between adjacent nodes is considered and modeled as a minimum weight sum problem (MWSP). To complete the entire data collection process, we performed path planning for AUVs and modeled it as an asymmetric traveling salesman problem (ATSP). A* algorithm and the Lin-Kernighan–Helsgaun (LKH) algorithm are designed to solve these problems, which are further nested in GA-NDQN to optimize the task allocation strategy for data collection. Finally, the effectiveness of the proposed scheme is verified by extensive simulation results.