PhD Proposal: Curriculum Learning for Hierarchical Task Networks
Ruoxi Li
Wednesday, December 21, 2022, 12:00-2:00 pm 
Abstract
Hierarchical Task Networks (HTNs) require a domain expert to provide knowledge
about the dynamics of the planning domain, including structural properties and
potential hierarchical problem-solving strategies. A significant knowledge engineering
burden for a domain expert is required to write HTN decomposition methods. Some
techniques overcome this burden, in part, by learning HTN methods after analyzing
the semantics of a solution trace for a sample planning problem.
My proposed research will focus on improving HTN learning using curricula. A
curriculum may improve an HTN learner by teaching it to learn simpler methods
before learning gradually more complex ones that make use of previously learned
methods. Such a curriculum may be generated from landmarks, each of which is a
fact that is true at some point in every plan that solves the problem.
In this dissertation proposal, first, I describe the completed algorithm HTNMakerC,
a modified version of HTN-Maker that learns HTN methods from a
curriculum, along with the theoretical analysis and discussion of the experiment
results. I propose to develop an algorithm that uses landmarks to generate curricula
that include interesting tasks to learn in each sample planning problem. Furthermore,
a more general curriculum learning technique would learn HTN methods from different
planning problems and their solution traces, and those problems are organized in a
way that the HTN methods learned from the previous problems could be utilized in
learning HTN methods from the later problems. Therefore, I propose to develop an
algorithm that performs curriculum learning among a sequence of sample problems.
In addition, I propose an algorithm that learns unlimitedly recursive HTN methods
that may increase the problem coverage of the set of learned methods. I also describe
my plan for the theoretical analysis of the proposed algorithms, their implementation,
and their evaluation in a variety of domains and problem sets from the International
Planning Competitions.
about the dynamics of the planning domain, including structural properties and
potential hierarchical problem-solving strategies. A significant knowledge engineering
burden for a domain expert is required to write HTN decomposition methods. Some
techniques overcome this burden, in part, by learning HTN methods after analyzing
the semantics of a solution trace for a sample planning problem.
My proposed research will focus on improving HTN learning using curricula. A
curriculum may improve an HTN learner by teaching it to learn simpler methods
before learning gradually more complex ones that make use of previously learned
methods. Such a curriculum may be generated from landmarks, each of which is a
fact that is true at some point in every plan that solves the problem.
In this dissertation proposal, first, I describe the completed algorithm HTNMakerC,
a modified version of HTN-Maker that learns HTN methods from a
curriculum, along with the theoretical analysis and discussion of the experiment
results. I propose to develop an algorithm that uses landmarks to generate curricula
that include interesting tasks to learn in each sample planning problem. Furthermore,
a more general curriculum learning technique would learn HTN methods from different
planning problems and their solution traces, and those problems are organized in a
way that the HTN methods learned from the previous problems could be utilized in
learning HTN methods from the later problems. Therefore, I propose to develop an
algorithm that performs curriculum learning among a sequence of sample problems.
In addition, I propose an algorithm that learns unlimitedly recursive HTN methods
that may increase the problem coverage of the set of learned methods. I also describe
my plan for the theoretical analysis of the proposed algorithms, their implementation,
and their evaluation in a variety of domains and problem sets from the International
Planning Competitions.
Examining Committee
Bio
Ruoxi Li is a PhD student working in AI planning advised by Professor Dana Nau. His research interests are in automated planning, hierarchical goal networks, planning and learning, and multi-agent systems.
This talk is organized by Tom Hurst