Planning as Heuristic Search: New Results

In the recent AIPS98 Planning Competition, the hsp planner , based on a forward state search and a suitable domain-independent heuristic, showed that heuristic search planners can be competitive with state of the art Graphplan and Satissability planners. hsp solved more problems than the other planners but often took more time or produced longer plans. The bottleneck in hsp is the computation of the heuristic for every new state. In this paper, we reformulate hsp so that this problem is avoided. The new planner, that we call hsp-r, is based on the same ideas and heuristic as hsp, but searches backward from the goal rather than forward from the initial state. We show that this change allows hsp-r to compute the heuristic only once. As a result, hsp-r can produce better plans, often in less time. For example, hsp-r solves each of the 30 logistics problems from Kautz and Selman in less than 3 seconds. This is two orders of magnitude faster than blackbox. At the same time, in almost all cases, the plans are substantially smaller. hsp-r is also more robust than hsp as it visits a larger number of states, makes deterministic decisions, and relies on a single adjustable parameter than can be safely xed for most domains. hsp-r, however, is not better than hsp accross all domains and in particular, in the blocks world, hsp-r fails on some large instances that hsp solves. We discuss also the relation between hsp-r and Graphplan, and argue that Graphplan is also best understood as a heuristic search planner with a precise heuristic function and search algorithm.