Learning Agents' Relations in Interactive Multiagent Dynamic Influence Diagrams

摘要

Solving interactive multiagent decision making problems is a challenging task since it needs to model how agents interact over time. Prom individual agents' perspective, interactive dynamic influence diagrams (I-DIDs) provide a general framework for sequential multiagent decision making in uncertain settings. Most of the current I-DID research focuses on the setting of n = 2 agents, which limits its general applications. This paper extends I-DIDs for n ＞ 2 agents, which as expected increases the solution complexity due to the model space of other agents in the extended I-DIDs. We exploit data of agents' interactions to discover their relations thereby reducing the model complexity. We show preliminary results of the proposed techniques in one problem domain.