A generate-and-test approach for computing “optimal” plans in SAT-based planning

Planning as Satisfiability (SAT) is the best approach for optimally (wrt makespan) solving classical planning problems. SAT-based planners, like SATPLAN, can thus return plans having minimal makespan guaranteed. However, the returned plan does not take into account plan quality issues introduced in the last two International Planning Competitions (IPCs): such issues include plans which minimize action costs and plans with “soft” goals, where a metric has to be optimized over actions/goals. Recently, an approach to address such issues has been presented, in the framework of planning as satisfiability with preferences: by imposing that one prefers “not to perform” actions, or “to satisfy” soft goals, the related system (called SATPLAN(P)) is guaranteed to return plans with minimal number of actions, or with maximal number of soft goals satisfied (thus in a qualitative case), and is extended to deal with the quantitative case. But, besides such feature, and the fact that the first computed plan is guaranteed to be “optimal”, it is well-known that introducing ordering in SAT heuristics can lead, at least theoretically, to significant degradation in performances: in SATPLAN(P), this phenomenon also happened experimentally on large planning problems with many preferences. In this paper, we present a different, generate-and-test, approach to tackle the problem of dealing with such optimization issues: without imposing any ordering, a (candidate optimal) plan is first generated, and then a constraint is added imposing that the new plan (if any) has to be “better” than the last computed, i.e., the plan quality is increased at each iteration. We implemented this idea in SATPLAN, and compared the resulting systems wrt SATPLAN(P) and SGPLAN on planning problems coming from IPCs. The analysis shows performance benefits for the new approach, in particular on planning problems with many preferences.