Generation of Hard Solvable Planning Problems

Existing models for sampling the space of all planning problems produce both solvable and unsolvable instances. In particular, when attempting to identify very hard ones, one encounters both, as the hardness in the standard models is accompanied by the possibility of an instance being solvable or unsolvable. This is a disadvantage of the standard models, because of the strong asymmetry of planning algorithms in handling solvable and unsolvable instances, and the preference to have solvable instances for the performance evaluation of planners. We look at the generation of hard problem instances that are guaranteed to have a plan. The basic method generates a state sequence which induces a valid plan. The main issue to be solved in this setting is the choice of the effects and the preconditions of the actions in the plan so that finding the plan becomes difficult. We provide solutions to this problem, and demonstrate its strength in terms of an experimental study with different types of planners. We argue that our methods are better suited to planner performance evaluation than the ones proposed by Bylander and Rintanen.