New MARL metric IEI boosts communication efficiency without performance loss

A new paper on arXiv (2606.07200) from researchers Xinren Zhang, Zixin Zhong, and Jiadong Yu introduces a principled metric called the Information Entropy Efficiency Index (IEI) for evaluating and training communication protocols in Multi-Agent Reinforcement Learning (MARL). Instead of chasing performance through ever-complex network architectures or expanding message sizes, IEI directly measures how much information entropy (a proxy for message richness) is required to achieve a given level of task success. The lower the IEI, the more efficient the protocol. By incorporating IEI into the training loss function, agents are incentivized to develop concise, information-dense messages that minimize overhead without sacrificing task performance.

Extensive experiments across diverse MARL benchmark tasks show that agents trained with the IEI objective match or outperform standard baselines while using significantly fewer bits or shorter messages. This result challenges the prevailing assumption that performance in cooperative multi-agent systems requires either complex communication or high bandwidth. Instead, the work demonstrates that focusing on information entropy efficiency can unlock scalable, high-performance multi-agent systems—critical for real-world applications like drone swarms, autonomous vehicle coordination, and distributed robotics where communication bandwidth is limited.