MARL and GNN foundations

We build the methodological foundations that make multi-agent learning work at scale. On the architectural side, we design graph neural networks that let agents reason over their teammates as a structured, variable-sized neighborhood rather than a flat observation — supporting generalization to team sizes and topologies unseen at training time. Our ModGNN framework and generalized f-mean aggregation formalize how information should be combined across these neighborhoods.

On the learning side, we develop MARL algorithms that tackle the exploration, credit assignment, and non-stationarity problems that make multi-agent learning hard, and we package them into BenchMARL, a standardized library used across the community to benchmark MARL methods under a common interface. A recurring theme is behavioral diversity: when and why identical agents benefit from learning distinct roles, and how to control that diversity to a target value rather than leaving it to chance.