Abstract: Graph neural networks (GNNs) have been used in various works for learning heuristics to guide
search for planning. However, they are hindered by their slow evaluation speed and their limited
expressiveness. It is also a known fact that the expressiveness of common GNNs is bounded by the
Weisfeiler-Lehman (WL) algorithm for testing graph isomorphism, with which one can generate
features for graphs. Thus, one may ask how do GNNs compare against machine learning models
operating on WL features of planning problems represented as graphs? Our experiments show that
linear models with WL features outpeform GNN models for learning heuristics for planning in the
learning track of the 2023 International Planning Competition (IPC). Most notably, our model
WL-GOOSE is the first model in the learning for planning literature which can reliably learn
heuristics from scratch that are competitive with $h^{\text{FF}}$ on problem sizes much larger than those
seen in the training set.
Submission Number: 4
Loading