Automatically Learning Hierarchical Task Network Methods from Planning Landmarks

Core Concepts

CURRICULAMA, an algorithm that automatically learns Hierarchical Task Network (HTN) methods from planning landmarks, eliminating the need for manual task annotations required by previous methods.

Abstract

CURRICULAMA is an algorithm that learns HTN methods from classical planning problems without requiring manual task annotations. It works by:

Extracting landmarks (facts that must appear in every solution plan) from the planning problem using landmark analysis.
Generating a curriculum of increasingly complex subtasks to learn, based on the landmark orderings.
Using the curriculum to guide the learning of HTN methods that can solve the original planning problem.

The key advantages of CURRICULAMA are:

It completely automates the HTN method learning process, eliminating the need for human-provided task annotations.
It uses landmarks to structure the learning process, starting with simpler subtasks and building up to more complex methods.
Experiments show CURRICULAMA has a similar convergence rate in learning a complete set of methods compared to the previous HTN-MAKER algorithm, while requiring no manual input.
CURRICULAMA first uses landmark analysis to extract a landmark graph from the classical planning problem. It then generates a curriculum of subtasks to learn, where each subtask corresponds to achieving a landmark. The curriculum starts with simpler subtasks and gradually increases in complexity.
CURRICULAMA then uses a subroutine called CURRICULEARN to learn HTN methods from this curriculum. CURRICULEARN analyzes plan traces to learn preconditions and subtasks for each method, reusing previously learned methods as subroutines when possible.
The paper proves that the methods learned by CURRICULAMA can be used to solve the equivalent hierarchical planning problem to the original classical planning problem. Experiments show CURRICULAMA learns methods comparably to the previous HTN-MAKER algorithm, while eliminating the need for manual task annotations.

Stats

The paper does not contain any explicit numerical data or statistics to extract.

Quotes

The paper does not contain any striking quotes to extract.

Key Insights Distilled From

Automatically Learning HTN Methods from Landmarks

by Ruoxi Li,Dan... at arxiv.org 04-10-2024

https://arxiv.org/pdf/2404.06325.pdf

Deeper Inquiries

How could CURRICULAMA's landmark ordering strategy be improved to further reduce the number of redundant methods it learns in some domains

To improve CURRICULAMA's landmark ordering strategy and reduce the number of redundant methods learned, several enhancements can be considered. Firstly, incorporating more sophisticated heuristics in the landmark ordering process could help prioritize landmarks more effectively. By analyzing the problem structure and dependencies more comprehensively, CURRICULAMA can generate more optimal landmark sequences, reducing the creation of unnecessary methods. Additionally, refining the reasonable ordering criteria to better capture the most efficient subgoal sequences could lead to a more streamlined learning process. By fine-tuning the landmark analysis algorithms and considering additional constraints, CURRICULAMA can minimize the generation of extraneous methods and improve overall efficiency.

What other types of structural knowledge, beyond HTN methods, could be learned using a curriculum-based approach guided by planning landmarks

Beyond HTN methods, a curriculum-based approach guided by planning landmarks could be applied to learn various types of structural knowledge in planning domains. For instance, task networks, action models, and state constraints could be learned using a similar curriculum learning framework. By leveraging landmarks to guide the learning process, algorithms could acquire knowledge about task dependencies, action preconditions, and state transitions. This approach could be extended to learn probabilistic models, temporal constraints, or resource allocation strategies in planning domains. By adapting the curriculum learning methodology to different types of structural knowledge, algorithms can efficiently acquire complex planning strategies from problem instances.

How might CURRICULAMA's performance scale as the complexity of the planning problems (e.g., number of objects, actions) increases, and what are the key factors that would impact its scalability

As the complexity of planning problems increases, CURRICULAMA's performance may be impacted by several key factors. The scalability of the algorithm could be influenced by the size of the planning domain, the number of objects, actions, and constraints involved, as well as the intricacy of the task dependencies. Additionally, the efficiency of landmark analysis and curriculum generation may be affected by the complexity of the problem structure and the diversity of landmarks present. The algorithm's scalability could also depend on the computational resources available, as larger problems may require more processing power and memory. By optimizing the landmark analysis algorithms, enhancing the curriculum learning strategies, and leveraging parallel processing capabilities, CURRICULAMA's performance can be scaled to handle increasingly complex planning scenarios.