Masked State Modeling and Intention Inference for Multi-Agent Communication
A representation-learning approach to inter-agent communication in cooperative multi-agent reinforcement learning that focuses on how a receiving agent assimilates messages, rather than only what is sent. Students learn how masked reconstruction of observation dimensions and prediction of teammates' joint actions build richer representations under partial observability, and how a meta-learned importance score selects which information to mask and share.
2501.00312
M2I2 is a multi-agent reinforcement learning framework that reframes the problem of integrating received messages at the receiving agent as a representation-learning task. Each agent performs masked …