Learning to Combine Admissible Heuristics Under Bounded Time

Usually, combining admissible heuristics for optimal search (e.g. by using their maximum) requires computing numerous heuristic estimates at each state. In many cases, the cost of computing these heuristic estimates outweighs the benefit in the reduced number of expanded states. If only state expansions are considered, this is a good option. However, if time is of the essence, we can do better than that. We propose a novel method that reduces the cost of combining admissible heuristics for optimal planning, while maintaining its benefits. We first describe a simplified model for deciding which heuristic is best to compute at each state. We then formulate an active learning approach to decide which heuristic to compute at each state, online, during search. The resulting technique, which we call selective max is evaluated empirically, and is shown to outperform each of the individual heuristics that were used, as well as their regular maximum, in terms of number of solved instances and average solution time.