Pursuit-Evasion Games for Multiple Unmanned Surface Vehicles Based on LD-MADDPG

This paper proposes a cooperative strategy based on the LD-MADDPG(Long Short-Term Memory-Differential Latency Gating Mechanism- Multi-Agent Deep Deterministic Policy Gradient) framework for the multi unmanned surface vessel(USV) pursuit game task. To address the problem of insufficient utilization of temporal information in individual decision-making and sparse group collaboration, Long Short-Term Memory(LSTM) is embedded in the network, enabling each USV to implicitly maintain a dynamic contextual memory, thereby making more continuous and robust maneuvering decisions in partially observable water surface environments. Further, to avoid the curse of dimensionality caused by the expansion of memory vectors over time, this paper introduces a differential saliency gating mechanism after the LSTM output: this module assigns importance weights to the implicit states at different historical moments, suppresses redundant information irrelevant to the current chase and escape phase, and retains only compact expressions sensitive to the critical situation of 'capture-escape'. This significantly enhances the efficiency of cluster collaboration. Simulation and measurement results show that the proposed algorithm has better convergence speed and average reward than the original Multi-Agent Deep Deterministic Policy Gradient(MADDPG).