log in  |  register  |  feedback?  |  help  |  web accessibility
PhD Defense: Automating Hierarchical Task Network Learning
Ruoxi Li
Monday, June 17, 2024, 10:00 am-12:00 pm
Abstract

Automating Hierarchical Task Network (HTN) learning is essential for reducing the knowledge engineering burden in automated planning systems. Traditional HTN learning techniques, like HTN-MAKER, face challenges related to scalability, efficacy, and the need for manual input. This dissertation fully automates HTN learning and significantly enhances the capabilities of the learned methods.


The proposed solution leverages curricula to learn simpler methods first and progressively tackling more complex ones. We use landmarks, facts that must be true in any plan solving the problem, as essential benchmarks to generate curricula. Additionally, the recognition of recursive task decomposition patterns allows for the learning of generalized methods, further improving the applicability of the learned methods.


The primary contributions of this dissertation include the development of CURRICULEARN, an algorithm that enhances HTN learning through the guidance of curricula; CURRICULAMA, an algorithm that automatically generates curricula from landmarks and utilizes CURRICULEARN to learn from those curricula; and METHODGENERALIZER, an algorithm that learns more generalized methods by capturing recursive task decomposition patterns. Some of these algorithms are theoretically analyzed and all of them are empirically evaluated across various domains, including those from the International Planning Competitions. CURRICULEARN is shown to learn fewer methods more efficiently compared to HTN-MAKER, resulting in higher planning success rates and reduced planning times. CURRICULAMA is proven to fully automate HTN learning while maintaining comparable performance compared to HTN-MAKER. METHODGENERALIZER further enhances the applicability of the learned methods, resulting in significantly high planning success rate for problems of various complexities.


In summary, this dissertation addresses the challenges in HTN learning by fully automating the HTN learning process, and significantly enhances the capabilities of the learned methods, contributing to the advancement of automated planning systems.

Bio

Ruoxi Li is a PhD student advised by Dr. Dana S. Nau. His research interests include AI Planning, Reinforcement Learning, Muti-agent System, and Robotics.

This talk is organized by Migo Gui