Passive and active decision postponement in plan generation

In "least-commitment" decision making, decisions are postponed until constraints force them to be made; any decision made when it is not forced is an "early commitment." Partial-Order Causal Link (POCL) planners typically postpone decisions about the ordering of plan steps, but rely to some degree on early commitments in making most other types of decisions. A prerequisite for the least-commitment approach is the ability to recognize when constraints force a decision to be made; most planners, however, take a "passive" approach to postponing most decisions, in which the relevant constraints are either not accessible or are not used by the planning algorithm. Part of this thesis looks at the inefficiencies that arise from passive postponement, and the extent to which these inefficiencies can be addressed by spending more time thinking about what to postpone. The Least-Cost Flaw Repair (LCFR) strategy (Joslin and Pollack, 1994) is one such response: postponed decisions are examined, and priority is given to decisions that are more tightly constrained. This approach increases planning efficiency, but is of limited effectiveness because interactions among postponed decisions are not recognized. Active postponement is a technique for allowing all constraints, including those that arise from postponed decisions and their interactions, to play a role in reasoning about the plan. Descartes, an algorithm that does active postponement, transforms planning problems into dynamic Constraint Satisfaction Problems (CSPs). That formulation leads to a natural identification of static sub-problems that can be solved by standard CSP algorithms. Active postponement also makes it possible to take a least-commitment approach to all planning decisions. Experimental results, however, show that least-commitment is not always best; sometimes "early commitment" does increase planning efficiency.