HIBCG: Information-Bottleneck Coordination Graphs for Multi-Agent RL

A recent study published on arXiv introduces Heterogeneous Information-Bottleneck Coordination Graphs (HIBCG) aimed at enhancing cooperative multi-agent reinforcement learning. This approach employs the graph information bottleneck (GIB) to establish a sparse coordination graph, where both the presence of edges and the capacity for messages are supported by theoretical foundations. HIBCG formulates a group-aligned block-diagonal prior that offers a clear criterion for edge retention, specifying which edges should be maintained and their density within each group. This innovation addresses the shortcomings of existing sparse-graph learners, which often depend on heuristic methods lacking formal assurances regarding topology and communication capacity distribution.