Searching with Probes: The Classical Planner PROBE

Heuristic search has been the mainstream approach in planning for more than a decade, with planners such as FF, FD, and LAMA being able to solve problems with hundreds of actions and variables in a few seconds (Hoffmann and Nebel 2001; Helmert 2006; Richter and Westphal 2010). The basic idea behind these planners is to search for plans using a search algorithm guided by heuristic estimators derived automatically from the problem (McDermott 1996; Bonet and Geffner 2001). State-of-the-art planners, however, go well beyond this idea, adding a number of techniques that are specific to planning. These techniques, such as helpful actions and landmarks (Hoffmann and Nebel 2001; Hoffmann, Porteous, and Sebastia 2004; Richter, Helmert, and Westphal 2008), are designed to exploit the propositional structure of planning problems; a structure that is absent in traditional heuristic search where states and heuristic evaluations are used as black boxes. Moreover, new search algorithms have been devised to make use of these techniques. FF, for example, triggers a best-first search when an incomplete but effective greedy search (enforced hill climbing) that uses helpful actions only, fails. In FD and LAMA, the use of helpful or preferred operators is not restricted to the first phase of the search, but to one of the open lists maintained in a multi-queue search algorithm. In both cases, dual search architectures that appeal either to two successive searches or to a single search with multiple open lists, are aimed at solving fast, large problems that are simple, without giving up completeness on problems that are not.