Planning via search in the input/output space

The computational complexity of planning can be reduced by the synthesis of an appropriate representational hierarchy. In order to provide a formal structure for the description of such a hierarchy, the concepts of temporal and spatial resolution are defined, and a technique for the automated construction of multilevel system representations is described. The operation of contraction (or focusing of attention) is introduced, and its use in accelerated, search-based planning within a hierarchical framework is proposed and discussed. The consistency and admissibility of plans generated using this approach are evaluated. It is shown why it is not feasible to search for a unique "best" plan. Instead, it is suggested that a minimax approach to randomized planning be implemented, which would attempt to minimize the penalty of eliminating the "best" solution from consideration.<
>