Relational GNNs Cannot Learn Features for Planning

Relational Graph Neural Networks (R-GNNs) are a GNN-based approach for learning value functions that can generalise to unseen problems from a given planning domain. R-GNNs were theoretically motivated by the well known connection between the expressive power of GNNs and , first-order logic with two variables and counting. In the context of planning, features refer to the set of formulae in with relations defined by the unary and binary predicates of a planning domain. Some planning domains exhibit optimal value functions that can be decomposed as arithmetic expressions of features. We show that, contrary to empirical results, R-GNNs cannot learn value functions defined by features. We also identify prior GNN architectures for planning that may better learn value functions defined by features.
View on arXiv@article{chen2025_2506.11721, title={ Relational GNNs Cannot Learn $C_2$ Features for Planning }, author={ Dillon Z. Chen }, journal={arXiv preprint arXiv:2506.11721}, year={ 2025 } }