Document Type



Master of Science


Computer Science

First Adviser

Munoz-Avila, Hector

Other advisers/committee members

Lopresti, Daniel P.


This thesis describes WORD2HTN, which is a novel and semantic approach for learning hierarchical task networks (HTN) and semantic division of goals from input plan traces. The semantic relationships are learned using machine learning to get the vector representations of the components of the plan trace. The semantic relationships are used to learn hierarchical landmarks, which in turn are used to make semantically divided HTNs. These learned HTNs can then be used for subsequent new problems in the domain that have a similar structure with the problems in the input plan traces. This work also improves the learning algorithm to include arithmetic conditions and effects. WORD2HTN was tested on 3 deterministic domains. These are Logistics or Transportation domain, Abstract Graph domain, and the Malmo interface for the Minecraft game. We show that WORD2HTN learns semantically divided HTNs. We also experimentally demonstrate that HTN planners using this have an exponential speedup in information-dense domains over the state of the art classical planner. Finally, we show that the HTNs learned in Minecraft can be used to achieve tasks faster with a cooperative agent controlled by the HTN planner’s output.