Emergent Compositional Concept Communication through Mutual Information in Multi-Agent Teams

Seth Karten, Siva Kailas, Katia P. Sycara

Abstract: In multi-agent reinforcement learning (MARL) with communication, coordination information (ordinal) is often required in addition to referential info about one's observations. The information bottleneck defines a trade-off between complexity and utility, which loses structure of latent information when compressed solely for utility. Thus, in this work, we use information theory to introduce information-rich, variational compositional communication to adequately embed referential information and to provide a contrastive objective to ground communication in intent-specific features without relying on reward. Each message is composed of a set of emergent concepts, which we show span the observations and intents. Messages are naturally compressed to the least number of bits.