We view planning as a consttaint-based composition problem: design a structure (i.e., a plan) by selecting and arranging elements (i.e., operators) from a pre-defined set of alternatives, such that the resultant structure satisfies certain constraints. We show how generalized operators allow for partial commitment to operator selec­tions. Using generalized operators gives the advantage of allowing constraints to be introduced and propagated without requiring an arbitrary choice to be made. The generalizations can be incrementally specialized as propagated constraints reduce the number of alternatives.

We describe OPIE, an implemented planner that uses incremental selection. We illustrate OPIE's performance on a small blocks world example. We also show how altering the control strategy allows OPIE to replicate the behavior of a variety of previous planning systems.