<ul><li>Value function decomposition methods for cooperative multi-agent reinforcement learning aim to compose joint values from individual per-agent utilities to ensure consistent action selection.</li><li>Existing methods like VDN and QMIX have limited representation capabilities, while QPLEX, the exception, is overly complex.</li><li>A new family of value function decomposition models called QFIX is introduced in this work, expanding representation capabilities with a fixing layer.</li><li>Empirical evaluation on multiple environments shows that QFIX enhances performance, learns stably, outperforms QPLEX, and uses simpler mixing models.</li></ul>

Fixing Incomplete Value Function Decomposition for Multi-Agent Reinforcement Learning

Discover more