Valerio Borelli, Alfonso Emilio Gerevini, Enrico Scala, Ivan Serina

Learning Heuristic Functions with Graph Neural Networks for Numeric Planning (Extended Abstract)

Introduction

Learning heuristic functions with graph neural networks for numeric planning (extended abstract). Learn how Graph Neural Networks (GNNs) derive effective heuristic functions for numeric planning. Our extended architecture optimizes best-first search with informative, efficient guidance.

Abstract

In this paper, we investigate the application of heuristics based on Graph Neural Networks (GNNs) to lifted numeric planning problems, an area that has been relatively unexplored. Building upon the GNN approach for learning general policies proposed by Staahlberg et al., we extend the architecture to make it sensitive to the numeric components inherent in the planning problems we address. We achieve this by observing that, although the state space of a numeric planning problem is infinite, the finite subgoal structure of the problem can be incorporated into the architecture, allowing for the construction of only a finite number of nodes. Instead of learning general policies, we train our models to function as a heuristic within a best-first search algorithm. We explore various configurations of this architecture and demonstrate that the resulting heuristics are highly informative and, in certain domains, offer a better trade-off between guidance and computational cost compared to other inductive and deductive heuristics.

Review

This extended abstract presents a timely investigation into the application of Graph Neural Networks (GNNs) for learning heuristic functions in the context of lifted numeric planning problems, an area acknowledged as relatively unexplored. Building upon established GNN approaches for policy learning, the authors propose a significant architectural extension to explicitly handle the numeric components inherent in such planning domains. The core motivation to develop informative heuristics for best-first search is well-aligned with critical needs in classical planning, where the efficiency and effectiveness of search are heavily reliant on good heuristic guidance. The technical contribution hinges on adapting GNNs to the peculiarities of numeric planning. While the state space is infinite, the authors cleverly address this challenge by leveraging the finite subgoal structure of a problem, which allows for the construction of a finite set of GNN nodes. This design choice is crucial for making GNNs applicable in this domain and demonstrates an insightful understanding of the problem's underlying structure. Instead of aiming for general policies, the model is specifically trained to provide heuristic estimates, a more focused objective that can yield significant practical benefits within a search algorithm. The exploration of various architectural configurations also suggests a thorough practical investigation. The preliminary results reported in this extended abstract are promising. The authors claim their GNN-based heuristics are "highly informative" and, in certain domains, demonstrate a superior trade-off between guidance quality and computational cost when compared against existing inductive and deductive heuristics. This claim, if substantiated in a full paper, would represent a notable advancement in the field of numeric planning. As an extended abstract, the details regarding the specific domains, experimental setup, and quantitative comparisons are understandably brief, but the proposed methodology and the initial findings certainly warrant further investigation and suggest a potentially impactful direction for future research in AI planning.

Full Text

Comments

You need to be logged in to post a comment.