State Super Sampling Soft Actor–Critic Algorithm for Multi-AUV Hunting in 3D Underwater Environment

Abstract

Reinforcement learning (RL) is known for its efficiency and practicality in single-agent planning, but it faces numerous challenges when applied to multi-agent scenarios. In this paper, a Super Sampling Info-GAN (SSIG) algorithm based on Generative Adversarial Networks (GANs) is proposed to address the problem of state instability in Multi-Agent Reinforcement Learning (MARL). The SSIG model allows a pair of GAN networks to analyze the previous state of dynamic system and predict the future state of consecutive state pairs. A multi-agent system (MAS) can deduce the complete state of all collaborating agents through SSIG. The proposed model has the potential to be employed in multi-autonomous underwater vehicle (multi-AUV) planning scenarios by combining it with the Soft Actor–Critic (SAC) algorithm. Hence, this paper presents State Super Sampling Soft Actor–Critic (S4AC), which is a new algorithm that combines the advantages of SSIG and SAC and can be applied to Multi-AUV hunting tasks. The simulation results demonstrate that the proposed algorithm has strong learning ability and adaptability and has a considerable success rate in hunting the evading target in multiple testing scenarios.