Contextualized Behavior Recommendation from Complex Agent-Based Simulations of Disasters

We present an approach for generating contextualized behavior recommendations from a large, data-driven, complex agent-based simulation. We extend a previous method for generating a summary description by decomposing the output of a simulation into a tree of causally-relevant states, and show how behavior recommendations can be generated by ranking these causally relevant states in terms of their impact on an outcome of interest. An end-user can provide a query specifying a partial state description, which is used to retrieve the appropriate set of states from the summary description. The structure of the tree is used to generate the contexts that differentiate the behavior recommendations. We apply our method to a very complex simulation of a disaster in a major urban area and present results for multiple queries.