Contextual Ranking of Behaviors for Large-scale Multiagent Simulations

As large-scale, complex multiagent simulations are becoming common, there is a need for new methods to analyze results of these simulations. One of the goals in such cases is to understand the effects of various behaviors on outcomes of interest. Here, we present a method for contextual ranking of behaviors where a partial context may already be provided in the query. Our approach uses causally-relevant states (states that have a measurable effect on the outcomes of interest), which provide the context for ranking behaviors. Apart from the partial context that may be provided in the query, our method also discovers any additional context that may affect behavioral ranking. We apply it to a large-scale disaster simulation and present results.