CP/SoCS/SAT 2025: Sylvie Thiebaux talk on Graph Learning for Planning

State-of-the-art methods for automated planning rely on heuristic state-space search. Prof. Thiebaux in her talk at CP/SoCS/SAT 2025 presented recent work on graph representation learning to guide the search of automated planners.  She introduced graph neural networks and other graph learning representations that exploit the relational structure of planning domains. They allow the planner GOOSE to learn heuristic cost estimates and state rankings from solutions to just a few small problems, and to solve substantially larger problems than those on which it was trained.

Perhaps surprisingly, her experimental results show that classical machine learning approaches vastly outperform deep learning ones in this context. Moreover, Greedy Best-First Search guided by the best learned heuristics rivals the state-of-the-art model-based planner, Lama, on the problems of the latest International Planning Competition Learning track.

This work opens the possibility that learned heuristics could soon become competitive replacements for existing model-based heuristics in automated planning, advancing both the theory and practice of the field.

You can explore the slides of the presentation here.