A Multi-Expert Intelligent Agent Reinforcement Learning Training Method Based on Cluster Confrontation
Yalei Niu, Xun Li, Ningyan Zhang, Yunhui Wang, Xinxin Zhang · 2025
This paper addresses typical red-blue swarm confrontation tasks by constructing a simulation environment based on Gym and developing a multi-expert agent reinforcement learning framework (MEARL). The core approach is to break down the adversarial scenarios into highlevel tasks and expert-level specialized tasks. Task-specific agents are designed to handle global coordination, while expert agents focus on domain-specific capabilities. Notably, these expert agents are pre-trained in typical scenarios to acquire specialized operational capabilities. When the number of intelligent agents of the red team and the blue team changes, the expert intelligent agent can adapt to the new scenario without re-training, thereby significantly reducing the computational cost and training time of the deep reinforcement learning system.