Conditional effects in Graphplan

TL;DR: The space of possible alternatives to factored expansion is described, and experimental results showing that factoredansion dominates full expansion on large problems are presented.

Abstract: Graphplan has attracted considerable interest because of its extremely high performance, but the algorithm's inability to handle action representations more expressive than STRIPS is a major limitation. In particular, extending Graphplan to handle conditional effects is a surprisingly subtle enterprise. In this paper, we describe the space of possible alternatives, and then concentrate on one particular approach we call factored expansion. Factored expansion splits an action with conditional effects into several new actions called components, one for each conditional effect. Because these action components are not independent, factored expansion complicates both the mutual exclusion and backward chaining phases of Graphplan. As compensation, factored expansion often produces dramatically smaller domain models than does the more obvious full-expansion into exclusive STRIPS actions. We present experimental results showing that factored expansion dominates full expansion on large problems.