Distributed Hierarchical Multi-Agent Reinforcement Learning for Real-World Adversarial Team Games
Wei Li, Bei Hui, Ke Lü · 2024
Real-world applications often require teams of agents to coordinate strategies against adversaries. However, existing MARL models overlook resource constraints and task complexity, limiting their practical use. To address this, we propose the Distributed Hierarchical Multi-agent Learning (DHML) framework, which uses distributed neural networks for efficient computation. A mean-field mechanism and an attentional communication module are further incorporated for scalability, and a hierarchical task decomposition is adopted based on the option-critic framework for handling complexity. Experiments on a simulation platform modeling real communication networks demonstrate DHML's effectiveness across various adversarial scenarios.