Learning Transition Statistics in Networks of Interacting Agents

Studying the decision-making of agents can reveal group behavior and internal lines of influence. We work with systems of interacting agents, where the decision-making of each agent is affected by their neighbors within some graph structure. As agents make choices, the stochastic transitions between chosen group actions can be learned, and thus the group behavior can be characterized and predicted. We express each element of the transition matrix P as a product of factors that depends on the agent neighborhood structure and leading to a separable estimator for the unknown p_ij of interest. This enables us to find a maximum likelihood estimator (MLE) for each factor and thus effectively estimate each p_ij with reduced complexity. We derive analytical concentration bounds for the error rates of this approach and demonstrate it on data sets.