Return to Tradition: Learning Reliable Heuristics with Classical Machine Learning

Dillon Ze Chen, Felipe Trevizan, Sylvie Thiebaux

Keywords: learning heuristics, classical machine learning, weisfeiler-lehman

TL;DR: learned heuristics with classical machine learning that outperform $h^{\text{FF}}$ on the learning track of the 2023 IPC, and theoretical connections to GNN and Description Logic features

Abstract:

There has been a renewed interest in applying machine learning to planning due to recent developments in deep neural networks, with a lot of focus being placed on learning domain-dependent heuristics. However, current approaches for learning heuristics have yet to achieve competitive performance against domain-independent heuristics in several domains, and have poor overall performance. In this work, we construct novel graph representations of lifted planning tasks and use the WL algorithm to generate features from them. These features are used with classical machine learning methods such as Support Vector Machines and Gaussian Processes, which are both fast to train and evaluate. Our novel approach, WL-GOOSE, reliably learns heuristics from scratch and outperforms the $h^{\text{FF}}$ heuristic. It also outperforms or ties with LAMA on 4 out of 10 domains. To our knowledge, the WL-GOOSE learned heuristics are the first to achieve these feats. Furthermore, we study the connections between our novel feature generation methods, previous theoretically flavoured learning architectures, and Description Logic features.

Primary Keywords: Learning

Category: Long

Student: Graduate

Submission Number: 54