Graphical Multiagent Models

Graphical models support compact representation of high-dimensional probability distributions when interaction among the variables exhibits locality. Graphical games (Kearns, et al.) exploit sparseness in an analogous way by factoring a payoff function based on local agent neighborhoods. Related formalisms such as MAIDs (Koller & Milch) extend graphical games to capture other structure in multiagent interactions, and NIDs (Gal & Pfeffer) carry the extension yet further to express models of non-optimizing behavior.

In the spirit of these and other prior works, we introduce Graphical Multiagent Models (GMMs), a straightforward representation for beliefs about multiagent outcomes using graphical models. GMMs provide a flexible way to express beliefs about the strategies agents will play, based on game-theoretic analysis, heuristic behavior models, observations of historical play, or other sources of knowledge. We show how GMMs support integration of multiple knowledge sources, leading to combined belief that outperforms belief based on single knowledge sources.

based on joint work with Quang Duong and Satinder Singh Baveja