Interactive dynamic influence diagrams (I-DIDs) provide an explicit way of modeling how a subject agent solves decision making problems in the presence of other agents in a common setting. To optimize its decisions, the subject agent needs to predict the other agents' behavior, that is generally obtained by solving their candidate models. This becomes extremely difficult since the model space may be rather large, and grows when the other agents act and observe over the time. A recent proposal for solving I-DIDs lies in a concept of value equivalence (VE) that shows potential advances on significantly reducing the model space. In this paper, we establish a principled framework to implement the VE techniques and propose an approximate method to compute VE of candidate models. The development offers ample opportunity of exploiting VE to further improve the scalability of I-DID solutions. We theoretically analyze properties of the approximate techniques and show empirical results in multiple problem domains.
|Title of host publication
|IJCAI'16: Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence
|AAAI Press/International Joint Conferences on Artificial Intelligence
|Number of pages
|Published - Jul 2016