Graph Learning for Planning: The Story Thus Far and Open Challenges (2412.02136v1)

Abstract: Graph learning is naturally well suited for use in planning due to its ability to exploit relational structures exhibited in planning domains and to take as input planning instances with arbitrary number of objects. In this paper, we study the usage of graph learning for planning thus far by studying the theoretical and empirical effects on learning and planning performance of (1) graph representations of planning tasks, (2) graph learning architectures, and (3) optimisation formulations for learning. Our studies accumulate in the GOOSE framework which learns domain knowledge from small planning tasks in order to scale up to much larger planning tasks. In this paper, we also highlight and propose the 5 open challenges in the general Learning for Planning field that we believe need to be addressed for advancing the state-of-the-art.

• The paper presents a comprehensive review where classical graph kernels outperform contemporary neural approaches in planning tasks.
• It introduces the GOOSE framework that scales planning via innovative graph representations and optimization formulations aligning state ranking with heuristic search.
• The study identifies open challenges including improving model expressivity, ensuring generalization across task sizes, and developing domain-specific optimization criteria.

Graph Learning for Planning: The Story Thus Far and Open Challenges

This paper provides a comprehensive review of the utilization of graph learning in planning domains, encapsulating the significant contributions made by the authors in the field. It underscores the potential of graph learning to enhance planning performance by leveraging the relational structures inherent in planning tasks. The authors focus on three pivotal components of graph learning for planning: graph representations, graph learning architectures, and optimization formulations. Through both theoretical exploration and experimental validation, the paper positions the GOOSE framework as a pivotal development in scaling planning from small to large tasks by learning domain knowledge.

Key Contributions

1. Graph Representations: The paper explores various graph representations for planning tasks and their expressive power when coupled with graph learning approaches. These representations are categorized into grounded graphs, lifted graphs with instantiation relations (IR), and lifted graphs with predicate relations (PR), each with distinct characteristics and applicability.
2. Graph Learning Architectures: Through an empirical comparison, the authors reveal that traditional machine learning models, specifically utilizing classical graph kernels, outperform contemporary neural network-based approaches like GNNs. This insight challenges the predominant reliance on deep learning approaches and suggests a paradigm shift towards classical models for specific planning contexts.
3. Optimization Formulations: The research proposes innovative formulations for optimizing graph learning models in planning. Unlike traditional methods that optimize based on cost-to-go estimates, the proposed optimizations frame heuristic functions as rankings of states, aligning better with the objectives of heuristic search strategies like GBFS.

Experimental Results

The authors conduct extensive experiments using benchmarks from the IPC23LT, focusing on classical and numeric planning scenarios. The results demonstrate the superior performance of classical machine learning models over deep learning models in symbolic planning, highlighting significant coverage gains in tasks of varying object quantities. Particularly in numeric planning, classical approaches provide competitive performance against state-of-the-art numeric planners, thus emphasizing their efficacy.

Open Challenges and Future Directions

The paper identifies several open challenges in the Learning for Planning (L4P) domain:

• Expressivity: There is a need to advance the expressive capacities of graph learning models, potentially incorporating recursion or leveraging inductive logic programming to capture complex planning tasks.
• Generalization: Understanding how planning models generalize across varying task sizes remains a critical area of exploration. Insights from complexity theory and planning-specific measures like novelty width may guide future research.
• Optimization Criteria: The paper stresses the importance of developing domain-specific optimization criteria. Current criteria may not universally apply across diverse planning landscapes, and future work should explore tailored formulations.
• Data Collection: Determining optimal strategies for data collection, akin to RL's exploration-exploitation tradeoff, poses a research challenge. Comprehensive strategies could significantly impact the efficacy of learned models.
• Benchmarking: Establishing standardized benchmarks and evaluation strategies for comparing diverse planning methodologies is crucial for advancing the field.

Conclusion

In sum, this paper delineates the state of graph learning for planning, highlighting groundbreaking frameworks and identifying crucial avenues for future investigation. By questioning existing paradigms and proposing new approaches, it sets the stage for continued advancement in AI planning, steering the field towards more robust, scalable, and efficient solutions.