Utilizing Problem Structure in Planning, A Local Search Approach

The book deals with general problem solving in the form given by the STRIPS planning formalism. Planning systems read in a STRIPS problem description, and then perform the search for a solution fully automatically, i.e., without any knowledge about the problem other than the transition rules that generate the search space. Starting from an existing approach to planning as heuristic search, I develop a novel system based on local search with a reened heuristic function. The system, named FF (short for Fast-Forward), achieves dramatic runtime improvements over all previous systems in a wide range of benchmark domains. A subsequent investigation determines the reasons for this performance: as it turns out, most of the benchmarks share certain kinds of problem structure. These structures imply certain characteristic qualities of FF's heuristic function, which in turn explain FF's behavior. I discuss the consequences of my ndings on ways of benchmarking in planning. 1 Planning Planning is a form of general problem solving. It has been one of the main topics of interest in AI from the outset. The task in the eld is to design problem solving techniques that can be applied to arbitrary problem domains and that thus model, or imitate, one of the distinguishing qualities of humans: the ability to reason about unseen domains. The overall goal in the eld can be formulated as making such general techniques competitive, within realistic application domains, with specialized solvers. General problems in this context are formulated by means of an initial state, a goal, and a set of actions. In the simplest and most widespread formalism, STRIPS, states are sets of propositional facts (those that are true in the state at hand). An action can be applied in a state if its precondition { a set of facts { is true in the state. When applying the action, its positive eeects (a set of facts) are added into the state, while its negative eeects (also a set of facts) are deleted. A plan is then a sequence of actions that transforms the initial state into a goal state { a state that satisses the goal (in the STRIPS case, a conjunction of facts). The benchmark problems traditionally used for evaluation in planning are simpliied formal adoptions of real-world problems, like e.g. transportation tasks (objects must be transported to diierent locations by means of vehicles) or construction tasks (complex objects must be assembled from their …