Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Efficiency Meets Symmetry Breaking

Yingbin Bai, Sylvie Thiebaux, Felipe Trevizan

TL;DR

This paper tackles the problem of symmetries in large planning search spaces by introducing Distincter, which combines a graph-based Typed Instance Learning Graph (TILG) with two pruning techniques to simultaneously boost learning efficiency and symmetry breaking. Action pruning uses automorphism orbits on TILG to eliminate symmetric actions, while state pruning leverages permutation-invariant GNN embeddings to hash and prune symmetric states during search. Integrated into Fast Downward, Distincter achieves historic coverage on the IPC Learning Track 2023, surpassing the traditional LAMA baseline in several domains and demonstrating robust gains through ablations. The work provides a practical pathway for scaling learning-based planners to large, symmetry-rich problems and suggests broader applicability to model-based planners as well.

Abstract

Learning-based planners leveraging Graph Neural Networks can learn search guidance applicable to large search spaces, yet their potential to address symmetries remains largely unexplored. In this paper, we introduce a graph representation of planning problems allying learning efficiency with the ability to detect symmetries, along with two pruning methods, action pruning and state pruning, designed to manage symmetries during search. The integration of these techniques into Fast Downward achieves a first-time success over LAMA on the latest IPC learning track dataset. Code is released at: https://github.com/bybeye/Distincter.

Learning Efficiency Meets Symmetry Breaking

TL;DR

This paper tackles the problem of symmetries in large planning search spaces by introducing Distincter, which combines a graph-based Typed Instance Learning Graph (TILG) with two pruning techniques to simultaneously boost learning efficiency and symmetry breaking. Action pruning uses automorphism orbits on TILG to eliminate symmetric actions, while state pruning leverages permutation-invariant GNN embeddings to hash and prune symmetric states during search. Integrated into Fast Downward, Distincter achieves historic coverage on the IPC Learning Track 2023, surpassing the traditional LAMA baseline in several domains and demonstrating robust gains through ablations. The work provides a practical pathway for scaling learning-based planners to large, symmetry-rich problems and suggests broader applicability to model-based planners as well.

Abstract

Learning-based planners leveraging Graph Neural Networks can learn search guidance applicable to large search spaces, yet their potential to address symmetries remains largely unexplored. In this paper, we introduce a graph representation of planning problems allying learning efficiency with the ability to detect symmetries, along with two pruning methods, action pruning and state pruning, designed to manage symmetries during search. The integration of these techniques into Fast Downward achieves a first-time success over LAMA on the latest IPC learning track dataset. Code is released at: https://github.com/bybeye/Distincter.
Paper Structure (24 sections, 3 theorems, 1 equation, 1 figure, 4 tables, 1 algorithm)

This paper contains 24 sections, 3 theorems, 1 equation, 1 figure, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and notation
  3. Distincter
  4. Typed Instance Learning Graph (TILG)
  5. Action Pruning
  6. State Pruning
  7. Experiments
  8. Datasets.
  9. Network structure.
  10. Training and evaluation.
  11. Results
  12. Ablation Study
  13. Conclusion
  14. Acknowledgments
  15. Proof of Theorem 1
  16. ...and 9 more sections

Key Result

Theorem 1

Let $A_i$ and $B_i$ be objects in ${\cal O}$ for $i \in \{1, \dots, n\}$ with $A_i \neq A_j$ and $B_i \neq B_j$ for all $i\neq j$. Let an action schema $\alpha \in \mathcal{A}$, and consider two ground actions, $a = \alpha(A_1, A_2, \dots, A_n)$ and $b = \alpha(B_1, B_2, \dots, B_n)$, applicable in

Figures (1)

  • Figure 1: A planning problem in the childsnack domain is illustrated using both ILG and TILG. Note that the colors used are for visualization purposes only, differing from the color encoding employed for computing automorphisms. Observe ILG lacks types (besides "Object") and static propositions.

Theorems & Definitions (7)

  • Definition 1
  • Definition 2
  • Theorem 1
  • Lemma 1
  • proof
  • Theorem 2
  • proof