Model identification in interactive influence diagrams using mutual information

Abstract

Interactive influence diagrams (I-IDs) offer a transparent and intuitive representation for the decision-making problem in multiagent settings. They ascribe procedural models such as influence diagrams and I-IDs to model the behavior of other agents. Procedural models offer the benefit of understanding how others arrive at their behaviors. Accurate behavioral models of others facilitate optimal decision-making in multiagent settings. However, identifying the true models of other agents is a challenging task. Given the assumption that the true model of the other agent lies within the set of models that we consider, we may utilize standard Bayesian learning to update the likelihood of each model given the observation histories of others' actions. However, as model spaces are often bounded, the true models of others may not be present in the model space. We then seek to identify models that are relevant to the observed behaviors of others and show how the agent may learn to identify these models. We evaluate the performance of our method on three repeated games and provide theoretical and empirical results in support.