Encoding HTN planning as a dynamic CSP

Abstract

Constraint satisfaction methodology has proven to be a successful technique for solving variety of combinatorial and optimization problems. Despite this fact, it was exploited very little in the planning domajn. In particular hierarchical task network planning (HTN) [2] seems to be suitable for use of constraint programming. The formulation of HTN planning problem involves a lot of structural information which can be used to prune the search space. Encoding of this structural information by means of constraint programming would provide an effective way for such pruning during the search for solution. This abstract describes a work currently in progress of which the goal is to develop a framework and techniques for solving HTN planning problems using constraint programming methodology. The first step to achieve the goal is to propose a suitable encoding of HTN planning problems into constraints. We encode HTN planning problem for a limited number of steps as a dynamic constraint satisfaction problem [2]. Our encoding translates each construct used in the formulation of HTN problem into a set of variables and constraints. The resulting constraint model is built hierarchically with global tasks (e.g. transport package from location A to location B) on the top of the hierarchy and with primitiv e actions (e.g. load package into the truck at location A) at the bottom. The dynamicity of our approach consists in construction of this hierarchical model during the search rfor solution. As the search proceeds and earlier decisions of the search algorithm become fixed, the model is extended with the parts modeling later decisions depending on the previous ones. Since our encoding relies on the intense combination of constraints via logical conjunctions, it is necessary to use a method for constraints combination that preserves stronger propagation. Such method is for example constructive disjunction [3]. © Springer-Verlag Berlin Heidelberg 2005.