Learning HTN Methods with Preference from HTN Planning Instances

Zhanhao Xiao, Hai Wan, Hankui Hankz Zhuo, Andreas Herzig, Laurent Perrussel, Peilin Chen

Abstract: The hierarchical task network (HTN) planning technique is used in a growing number of real-world applications. However in many domains, such as the logistics domain, as there exist thousands of cases, it is difficult and time-consuming for humans to specify all HTN methods to cover all desirable plans. This suggests that it is important to learn HTN methods to accomplish the tasks via decomposition. The traditional HTN-method learning approaches require complete executable plans and annotated tasks, which are often difficult to acquire in real-world applications. In this paper, we propose a novel framework to learn HTN methods from HTN instances with incomplete method sets and without annotated tasks. Besides, previous approaches demand total orders on the subtasks in the methods while our approach is capable of learning methods with partial orders. To reduce the number of methods learned, we consider priorities on methods and compute the minimal set of methods based on prioritized preferences. By taking experiments on three well-known planning domains, we demonstrate that our approach is effective, especially on solving new HTN problems.