Learning Hierarchical Task Knowledge for Planning

Pat LangleyAAAI 2025 · DOI

TL;DR

This paper explores how AI systems can automatically learn hierarchical task knowledge, similar to how humans break down complex activities. By learning these structures, AI planners can become significantly more efficient and scalable, avoiding the tedious manual effort currently required to encode such knowledge. The authors identify key challenges and propose future research directions to advance this crucial area of AI.

What Problem Does This Paper Address?

Modern AI planning systems often struggle with complex tasks that require breaking them down into smaller, manageable subproblems. While Hierarchical Task Networks (HTNs) offer a powerful way to represent this kind of knowledge, manually creating these structures is time-consuming, expensive, and error-prone, much like early expert systems. This paper addresses the challenge of automatically acquiring HTN knowledge from experience, drawing parallels to the success of machine learning in other AI domains. The goal is to enable AI planners to learn how to decompose problems efficiently, thereby reducing search space and improving performance, especially for novel or complex scenarios.

Methods

The paper reviews existing approaches to learning plan expertise, categorizing them into three main paradigms:

  • Acquiring search-control knowledge: Learning rules to guide the planner's search by selecting, rejecting, or preferring states, goals, or operators.
  • Forming macro-operators: Learning sequences of actions that act as single, more powerful operators, allowing planners to take larger steps.
  • Learning hierarchical task networks (HTNs): The paper's central focus, where expertise is encoded as hierarchical methods that break down tasks into subtasks with specific conditions for application.

For HTN learning specifically, the paper outlines three core subproblems:

  1. Identifying hierarchical structures: Discovering the nested relationships between tasks and subtasks within a plan.
  2. Unifying method heads: Determining when different task decompositions should share the same high-level task name or "head," allowing for disjunctive or recursive knowledge.
  3. Finding method conditions: Learning the specific circumstances under which a particular task decomposition should be applied.

The paper discusses both analytical (reasoning over plan steps) and inductive (finding regularities in sample solutions) approaches to tackle these subproblems.

Datasets

Not discussed in the paper. The paper focuses on the forms of input for HTN learning, which include domain predicates, domain operators, and sample traces of problem solutions (generated by the planner or provided by a human).

Key Insights and Ideas

  • Human-like Decomposition: The core idea is that AI can learn to decompose complex tasks in a way analogous to how humans naturally break down everyday activities, leading to more intuitive and efficient planning.
  • Learning vs. Manual Construction: The paper strongly advocates for learning HTNs over manual construction, highlighting the significant advantages in terms of cost, maintenance, and error reduction, mirroring the broader impact of machine learning.
  • HTNs as a Powerful Abstraction: HTNs provide a natural way to represent knowledge at multiple levels of abstraction, making plans more interpretable and enabling significant reductions in search complexity.
  • The "Weakest Link": Identifying the conditions under which methods should be applied is identified as the most challenging aspect of HTN learning, requiring further research.
  • Beyond Planning: Hierarchical knowledge isn't just for plan generation; it can also be used for knowledge-guided reactive control and plan recognition.

Key Results

The paper is primarily a survey and a call for future research, rather than a presentation of novel experimental results. It reviews existing work and identifies challenges. The "results" are more conceptual:

  • Demonstrating the theoretical and empirical benefits of domain knowledge for planning, especially for HTNs.
  • Outlining the necessary inputs and subproblems for HTN learning.
  • Identifying seven key challenges that need to be addressed for HTN learning to reach its full potential.

Limitations and Future Work

The paper explicitly lists seven challenges for future work:

  1. Improving condition finding: Developing faster and more human-like methods for learning method conditions.
  2. Acquiring binding constraints: Learning to specify desirable bindings for variables within methods, addressing limitations of current approaches.
  3. Integration with action learning: Combining HTN learning with learning action models from observation and experimentation to automate the process more fully.
  4. Learning from partial solutions: Developing methods that can learn from incomplete or incorrect sample solutions, as human-provided examples may not always be perfect.
  5. Forming conceptual predicates: Automating the creation of new relational predicates that describe states abstractly, reducing reliance on manual state abstraction.
  6. Cumulative acquisition of methods: Enabling systems to learn and build upon methods over long periods in complex domains, leading to more robust knowledge bases.
  7. Evaluating candidate methods: Developing mechanisms to detect and address the "utility problem" (where learned methods might not actually improve performance) and to evaluate methods based on their track record.

Summary updated

BibTeX
@inproceedings{langley2025learning,
  author = {Pat Langley},
  title = {{Learning Hierarchical Task Knowledge for Planning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35091},
}
Figure