Finding State Similarities for Faster Planning

In many planning applications one can find actions with overlapping effects. If for optimally reaching the goal all that matters is within this overlap, there is no need to consider all these actions – for the task at hand they are equivalent. Using this structure for speed-up has previously been proposed in the context of least commitment planning. Of a similar spirit is the approach for improving best-first search based planning we present here: intuitively, given a set of start states, reachable from the initial state, we plan in parallel for all of them, exploiting the similarities between them to gain computational savings. Since the similarity of two states is problem specific, we explicitly infer it by regressing all relevant entities, goal, heuristic function, action preconditions and costs, over the action sequences considered in planning. If the resulting formulae mention only fluents whose values the two states have in common, it suffices to evaluate the formulae in one of them. This leads to computational savings over conventional best-first search.