Planning by Local Search (position Paper)

Solving planning problems by local search provides a promising alternative to both classical (i.e., purely deductive) AI planning and the use of special purpose planning algorithms. The success of this approach has been manifested by the recent award-winning paper Kautz and Selman, 1996], which proves that a speciic local search method can substantially outperform the best existing planning procedures. This particular approach is of course only one out of many possible ways to do planning locally. Ongoing research in both reasoning about actions and local search paradigms gives rise to a whole variety of roads worthwhile to explore. In this note, we name the most important principles underlying the approach of Kautz and Selman, 1996]. For each we propose alternatives and argue why there is hope that following these suggestions leads to even more impressive results. The underlying action calculus. The method described in Kautz and Selman, 1996] is based on axiomatizing planning problems using the representation technique of Situation Calculus. As a consequence, the approach has to cope with the Frame Problem, in particular with the inferential aspect thereof. That is to say, one-by-one all properties must be carried to the next situation which are unchanged when an action is performed. Accordingly, the resulting axiomatization includes extra clauses which serve precisely this purpose. This favors the employment of alternative action calculi that provide better solutions to the inferential aspect of the Frame Problem. We pursue the use of the Fluent Calculus paradigm (see, e.g., Thielscher, 1997]) as a promising candidate. The basic concept there is to encode states as terms so that the eeects of actions can be modeled by removal and addition of sub-terms. Consequently, no eeort is needed to conclude that unaaected properties continue to hold|the corresponding sub-terms being untouched induces this automatically. The underlying ontology. The ontology on which Kautz and Selman, 1996] is based is Strips-like but additionally supports the notion of state constraints. The latter, however, is not exploited for the automatic proliferation of indirect eeects of actions. Rather these eeects are all compiled into the action descriptions. As a consequence, a single cause-eeect relation may be multiply encoded, which gives rise to possibly noticeable redundancy. The problem of how to derive the correct indirect eeects from underlying state constraints is known under the name Ramiication Problem. Recently it has been shown that the incorporation of a suitable notion of causality provides a broadly …