Matching Problem Features with Task Selection for Better Performance in HTN Planning

During the planning process, a planner may often have many different options for what kind of plan refinement to perform next (for example, what task or goal to work on next, what operator or method to use to achieve the task or goal, or how to resolve a conflict or enforce some constraint in the plan). The planner’s efficiency depends greatly on how well it chooses among these options. In this paper, we present and compare two types of strategies that an HTN planner may use to select which task to decompose next. Both strategies facilitate efficient planning by making it easier for the planner to identify plans that can be pruned from the search space--but since the strategies accomplish this in two different ways, each works better on different kinds of problems. We present experimental results showing how characteristics of the planning domain can be used to predict which strategy will work best, so that these domain characteristics can be used to select strategies across application domains when building practical planning systems.