Make Use of Your Search Effort: Data Augmentation for Learning Effective Planning Heuristics from Optimal Plans

Abstract

Learning planning heuristics to guide search has proven to be an effective way to improve satisficing planning with data, while preserving the soundness and completeness guarantees of the planner. Prior work has shown that learning from optimal plans can yield more effective heuristics, since optimal plans provide high-quality supervision about promising search directions. However, generating training data is computationally expensive, as it requires solving challenging planning tasks optimally. We therefore propose search-free data augmentation methods that can substantially increase the size of training data from existing optimal plans and corresponding search traces, without solving any additional planning problems. In particular, we consider three complementary data augmentations with optimality guarantees on the augmented data: (i) Intermediate Goals, which converts subplans of an optimal plan into new optimal plans to alternative, intermediate goal states; (ii) Closed-List Trajectories, which leverages the fact that under A* with a consistent heuristic each expanded state is reached by an optimal path; and (iii) Partial-Order Plan, which samples multiple linearizations from the underlying POP, which are also optimal plans under state-independent action costs. Experiments on the IPC 2023 Learning Track demonstrate that our data augmentations significantly improve planning performance across many domains for two state-of-the-art frameworks, GOOSE (WLF+ranking) and Distincter (GNN+regression). For the first time, our data augmentation techniques enable GOOSE to surpass the coverage achieved by the first iteration of the LAMA planner.