Explanations of Empirically Derived Reactive Plans

Given an adequate simulation model of the task environment and payoff function that measures the quality of partially successful plans, competition-based heuristics such as genetic algorithms can develop high performance reactive rules for interesting sequential decision tasks. We have previously described an implemented system, called SAMUEL, for learning reactive plans and have shown that the system can successfully learn rules for a laboratory scale tactical problem. In this paper, we describe a method for deriving explanations to justify the success of such empirically derived rule sets. The method consists of inferring plausible subgoals and then explaining how the reactive rules trigger a sequence of actions (i.e., a strategy) to satisfy the subgoals.