A model recognition approach for solving interactive dynamic influence diagrams

The interactive dynamic influence diagrams provides a way to model and solve multi-agent decision-making problems from the perspective of the subject agent. The subject agent usually optimizes its own decisions by predicting the behavior of other agents. The exponential increase in the model of other agents over time bring great difficulties to the decision. In this paper, we propose a learning algorithm based on dynamic Bayesian network, and utilize standard Bayesian learning to update the likelihood of each candidate model given the observation histories. By excluding unrelated or weakly related models, the model space is fundamentally compressed. The experimental results show the advantages of the algorithm: on the one hand, the method successfully determines the real model of other agents to predict their behavior; on the other hand, compared with other methods, the method gets a higher return.