From Task Definitions and Plan Traces to HTN Methods

Hierarchical Task Network (HTN) planning is an important, frequently studied research topic in artificial intelligence. Researchers have reported work on its formalisms and applications (Erol, Hendler, & Nau 1994; Smith, Nau, & Erol 1998; Nau et al. 2005). In HTN planning, complex tasks are decomposed into simpler tasks until a sequence of primitive actions is generated. HTN planning is frequently studied because it is analogous to a common model of human thought and because it has allowed impressive gains in execution time when compared to classical planners. Despite these advantages, a major hurdle for the use of HTN planning is the need for an HTN domain description. In fact, a controversy in the AI planning research community surrounds the recent efficiency gains obtained with HTN planning because the domain descriptions sketch the underpinnings of the solutions. Therefore, it has been argued that a significant knowledge engineering effort is required to obtain such domain descriptions. A domain description is a collection of knowledge constructs describing the target domain. In HTN planning, a domain description consists of the action model and the task model. The action model encodes knowledge about valid actions or primitive tasks changing the world state. The task model encodes knowledge about how to decompose tasks into subtasks, and is the part of the domain description that has been argued to be difficult to obtain. Given the large interest in HTN planning, it is surprising that little research has been done on learning task models. The bulk of research involving planning and learning has focused on search control knowledge (Zimmerman & Kambhampati 2003). We present HTN-MAKER (Hierarchical Task Networks with Minimal Additional Knowledge Engineering Required), an offline and incremental algorithm for learning task models. HTN-MAKER receives as input a collection of plans generated by a STRIPS planner, an action model, and a collection of task definitions, and it produces a task model. When combined with the action model, this task model results in an HTN domain model that may be used by an HTN planner to solve problems in the domain. An extended version of this paper has been submitted to the Planning and Learning workshop at ICAPS-07, which includes a description of the learning algorithm and empirical evaluation. Related Research