Enhancing Multi-Agent Reinforcement Learning via Entity Aware-Interaction Network
Xinyu Shen · 2024
This paper addresses a significant challenge in multi-agent reinforcement learning (MARL): the paradoxical underperformance of theoretically superior algorithms like QTRAN compared to simpler ones such as QMIX and VDN in practical applications. We identify the root cause as inefficient network representations in QTRAN implementations. To resolve this, we propose a novel, high-efficiency network structure incorporating relational graph network and attention graph network modules. This structure significantly enhances QTRAN's sample efficiency and overall performance. By conceptualizing multi-agent systems as graph structures, our approach enables more effective representation of agent interactions and relationships. Experimental results demonstrate that our enhanced QTRAN outperforms state-of-the-art algorithms like QMIX and VDN in benchmark environments. This work contributes to bridging the gap between theoretical advancements and practical performance in MARL, particularly in fully cooperative tasks.