93 / 2025-04-14 19:04:23

Multi-Agent Fault-Tolerant Control based on Adaptive Disturbance Suppression in the Reinforcement Learning State Space

MARL, fault-tolerant control, state space, mutual information

Multi-agent systems are becoming increasingly critical in practical manufacturing, transportation, and warehousing applications. State space-based multi-agent reinforcement learning (MARL) provides comprehensive multi-agent control strategies for complex environments. However, state space chaos caused by sensor failure-induced disturbances results in the degradation of the MARL control performance. To mitigate state space chaos, we propose an MARL-based adaptive state disturbance suppression method (ADSDS). First, we develop a contrastive encoder in the actor network to adaptively select reward-guided features for fault/redundant observation exclusion. Second, we design a regularized robust control based on prioritized replay and adaptive gradient regularization. Finally, we incorporate uncertain fault conditions and dynamic environments to validate robust control performances of the proposed ADSDS. The experiments demonstrate that our method has faster convergence and higher cumulative returns than existing methods in faulty environments.