Conditional Effects in Graphplan

Graphplanhas attracted considerable interest because of its extremelyhigh performance, but the algorithm’s inability to handle action representations moreexpressive than STRIPSis a major limitation. In particular, extending Graphplanto 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 newactions called components, one for each conditional effect. Because these action components are not independent, factored expansion complicates both the mutual exclusion and backwardchaining phases of Graphplan. As compensation, factored expansion often produces dramatically smaller domainmodels than does the more obvious full-expansion into exclusive STRIPSactions. Wepresent experimental results showingthat factored expansion dominatesfull expansion on large problems.

Authors

• Corin R. Anderson
• David E. Smith 0001
• Daniel S. Weld

Keywords

No keywords are indexed for this paper.