Characterizing and Automatically Finding Primary Effects in Planning

The use of primary effects of operators in planning is an effective approach to reduce search costs. However, the characterization of 'good" primary effects has remained at an informal level. In this paper we present a formal criterion for selecting useful primary effects, which guarantees planning efficiency, completeness, and optimality. We also describe an inductive learning algorithm based on this criterion that automatically selects primary effects of opera- tors. Both the sample complexity and the time complexity of our learning algorithm are polynomial in the size of the domain.