An implementationan evaluation of a hierarchical nonlinear planner

Abstract

Planning is the process of generating sequences of actions in order to provide a method for agents to modify the state of the world in which they exist. It is well known that the application of abstraction can greatly reduce the search involved in creating plans. In this talk, an empirical evaluation of several different approaches to reducing search in AbTweak, an abstract nonlinear planning system, are presented.

To solve a complex problem using abstraction, one would like to protect parts of the abstract solution that have already been completed during the abstract plan refinement. However, it has not been well understood how this protection can be done profitably; some subgoal protection may indeed result in a decrease in efficiency. The Monotonic Property has been proposed as a form of goal protection in abstract planning. Different versions of this property are investigated with respect to their effect on planning performance in several domains. Furthermore, a series of empirical tests of the utility of the properties are presented, along with an evaluation of different search strategies for abstract planning.

In addition, we present a novel abstract planning control strategy known as LeftWedge. This strategy adds a depth-first flavour to a complete search strategy by taking advantage of the fact that less abstract solutions are generally closer to a concrete level solution than more abstract ones. The relative benefit of this approach under various criticality assignments is demonstrated through comparisons with breadth-first strategy. Furthermore, two subgoal selection strategies are presented and compared empirically.