Role-cognitive decomposition for multi-agent reinforcement learning in StarCraft II micromanagement

Abstract

In complex, partially observable environments such as real-time strategy games, relying on decentralized multi-agent systems to learn efficient cooperative strategies is a challenging task. This paper proposes a Role Cognition Decomposition algorithm based on Multi-Agent Reinforcement Learning (RCD-MARL), which combines the ideas of role discovery and cognitive-level value decomposition. This algorithm infers agents' roles based on their potential influence and local observation dynamics, forming a specialized role cognition strategy. On this basis, a monotonic mixture network is adopted to decompose the global Q-value according to these roles, ensuring the consistency of team strategies. We benchmarked RCD-MARL against advanced baselines, including role-based methods such as QMIX and RoMIX, on StarCraft II micro-operation tasks using the SMAC and improved VLM-Attention environments. Experimental results show that the algorithm achieves good performance, with an average win rate improvement of 15-25% and a convergence speed increase of 30-40% compared to other baselines.