Leveraging Action Relational Structures for Integrated Learning and Planning

TL;DR

This work addresses the gap between learning-based heuristics and search strategies in classical planning by introducing partial-space search, which leverages the hierarchical structure of PDDL action schemas to refine the search space. It defines action set heuristics that operate on state–partial-action pairs, and provides two graph representations (AOAG and AEG) to learn these heuristics, either by restriction from state-space heuristics or from scratch via learning. The authors implement a new planner, LazyLifted, and demonstrate that partial-space search with learned action-set heuristics outperforms state-of-the-art learned heuristics on IPC 2023 LT and high-branching-factor benchmarks, and even surpasses LAMA in combined evaluations. The approach also yields richer training data and yields more informed heuristics, highlighting the practical impact of integrating learning more deeply into the search process for planning domains with high branching factors and complex action schemas.

Abstract

Recent advances in planning have explored using learning methods to help planning. However, little attention has been given to adapting search algorithms to work better with learning systems. In this paper, we introduce partial-space search, a new search space for classical planning that leverages the relational structure of actions given by PDDL action schemas -- a structure overlooked by traditional planning approaches. Partial-space search provides a more granular view of the search space and allows earlier pruning of poor actions compared to state-space search. To guide partial-space search, we introduce action set heuristics that evaluate sets of actions in a state. We describe how to automatically convert existing heuristics into action set heuristics. We also train action set heuristics from scratch using large training datasets from partial-space search. Our new planner, LazyLifted, exploits our better integrated search and learning heuristics and outperforms the state-of-the-art ML-based heuristic on IPC 2023 learning track (LT) benchmarks. We also show the efficiency of LazyLifted on high-branching factor tasks and show that it surpasses LAMA in the combined IPC 2023 LT and high-branching factor benchmarks.

Figures (7)

• Figure 1: Partial action tree for a Blocksworld task with objects $a$ and $b$. Darker node colours indicate higher specificity.
• Figure 2: Example expansion trees of partial-space search (left) and state space search (right). Solid lines indicate expanded nodes, dashed lines indicate unexpanded nodes. Expansion decisions are guided by an informed heuristic.
• Figure 3: AOAG (Def. \ref{['def:aoag']}) of a Blocksworld Instance with blocks $a$ and $c$ on the table and $b$ being held, here the goal is to place $b$ on $a$ and $\Lambda$ contains all applicable $\mathtt{stack}$ actions.
• Figure 4: AEG (Def. \ref{['def:aeg']}) of a Blocksworld Instance with blocks $a$ and $c$ on the table and $b$ being held, with the goal to place $b$ on $a$ and $\Lambda$ containing all applicable $\mathtt{stack}$ actions.
• Figure 5: Comparison of partial-space search versus state-space search on branching factor (top, measured as average number of successors per expansion), expansions (middle), and evaluations (bottom), for AOAG and AEG. The $x=y$ line is shown in the diagonals. Points below it favour partial-space search and above it favour state-space search.

Theorems & Definitions (8)

• Definition 1
• Definition 2
• Theorem 1
• proof
• Definition 3
• Definition 4: Restriction heuristic
• Definition 5: AOAG
• Definition 6: AEG