Making Graphplan Goal-Directed

Abstract

The Graphplan algorithm exemplifies the speed-up achieveable with disjunctive representations that leverage opposing directions of refinement. It is in this spirit that we introduce Bsr-graphplan, a work in progress intended to address issues of scalability and expressiveness which are problematic for Graphplan. Specifically, we want to endow the planner with intelligent backtracking and full quantification of action schemata. Since Graphplan employs a backward chaining search, it lacks the necessary state information to efficiently support these mechanisms. We hypothesize that alternatively pointing the search in the direction of the goals provides an sufficient amelioration. Further, we demonstrate that a forward chaining search strategy can be competitive by enforcing ordering constraints on the developing plan prefix. This is accomplished by using operators of a plangraph constructed in a top-down fashion to extend the plan prefix, and by introducing an additional data structure – a constraint tree – which is constructed by regressing subgoal information through the graph operators.