This paper extends the framework of dynamic influence diagrams (DIDs) to the multi-agent setting. DIDs are computational representations of the Partially Observable Markov Decision Processes (POMDP), which are frameworks for sequential decision-making in single agent settings. The Interactive Dynamic Influence Diagrams (I-DIDs), presented here, are computational representations of Interactive Partially Observable Markov Decision Processes (I-POMDPS). I-POMDPs generalize POMDPs to multi-agent settings by including the models of other agents in the state space. In I-DIDs agents maintains their beliefs over models of other agents. They then use these models to predict the other agents' likely behavior and compute their own best response given these predications. Models of other agents could themselves be I-DIDs, DIDs, or simply probability distributions over their actions. The possibility that models are I-DIDs leads to recursive nesting of models. To ensure that models are always c...
Kyle Polich, Piotr J. Gmytrasiewicz