Role Discovery in Observed Multi-Agent Systems Over Time through Matrix Factorization

Brian Reily, Michael Don, John G. Rogers, Christopher M. Reardon

Abstract: Understanding unknown multi-agent systems solely from observations without prior knowledge of the systems' composition or structure is critical for effectively responding to and interacting with them. Whether the unknown system consists of humans, robots, or other entities, the capability to discover the roles that various agents play in the multi-agent system is necessary to fully understand it. While existing work often focuses on predicting future trajectories or behaviors, there has been little research on identifying agents that share roles within a multi-agent system. Discovering shared roles enables a fuller understanding of the system and its future behavior, i.e., agents that share a role could be expected to behave similarly. In this paper, we propose a novel approach for role discovery in an observed multi-agent system. We first present a method to learn a unified temporal representation of the multi-agent system through a temporally weighted approximation of graphs describing relationships between agents at each time step. We then present our main contribution, where we formulate role discovery as a regularized optimization problem with the goal of learning the optimal role assignment based on the unified temporal representation. Our approach learns probabilities that agents play different roles while also discovering the number of distinct roles that exist in the multiagent system, and is proven to converge to the optimal solution. We also introduce a new role recognition dataset and evaluate on an existing dataset, showing that our approach outperforms existing methods in discovering roles in an observed multi-agent system.