Graphical Multiagent Models

I introduce Graphical Multiagent Models (GMMs): probabilistic graphical models that capture agent interactions in factored representations for efficient inference about agent behaviors. The graphical model formalism exploits locality of agent interactions to sup- port compact expression, and provides a repertoire of inference algorithms for efficient reasoning about joint behavior. I demonstrate that the GMM representation is sufficiently flexible to accommodate diverse sources of knowledge about agent behavior, and to com- bine these for improved predictions. History-dependent GMMs (hGMMs) extend the static form of GMMs to support representation and reasoning about multiagent behavior over time, offering the particular advantage of capturing joint dependencies induced by abstrac- tion of history information. Computational experiments demonstrate the benefits of explic- itly modeling joint behavior and allowing expression of action dependencies when given limited history information, compared to alternative models.
Many multiagent modeling tasks that employ probabilistic models entail constructing dependence graph structures from observational data. In the context of graphical games where agents' payoffs depend only on actions of agents in their local neighborhoods, I for- mally describe the problem of learning payoff dependence structures from limited payoff
x
observations, and investigate several learning algorithms based on minimizing empirical loss. I show that similar algorithms can also be applied to learning both dependence struc- tures and parameters for hGMMs from time-series data. I use data on human-subject ex- periments in evaluating performance of learned models for a consensus dynamics scenario. Analysis of learned graphical structures reveals patterns of action dependence not directly reflected in observed interactions. The problem of modeling information diffusion on par- tially observed networks further provides another demonstration of hGMM flexibility, as unobserved edges may induce correlations among node states. I find that learning graphical multiagent models for a given network structure, which can compensate for induced cor- relations, from diffusion traces outperforms directly learning the missing edges from the same data source.