New AI method improves multi-agent coordination by learning sparse communication graphs from historical data rather than single observations

Researchers propose Latent Temporal Sparse Coordination Graph (LTS-CG) to better represent how multiple AI agents should cooperate in reinforcement learning tasks

The approach uses agents' historical observations to build an agent-pair probability matrix, reducing reliance on one-step observations that miss important context

Sparse graph sampling reduces computational complexity and improves scalability compared to dense graph methods that require expensive action-pair calculations

LTS-CG simultaneously captures agent dependencies and accounts for uncertainty in agent relationships, enabling more efficient knowledge exchange between agents