wlplan 1.0.1

WLPlan: Relational Features for PDDL Planning

WLPlan

WLPlan is a package for generating embeddings of PDDL planning problems for machine learning tasks. It supports both classical and numeric planning problems.

Installation

Python Interface

The Python interface can be installed simply with

sh install.sh

C++ Interface

The C++ interface can be installed in your project by running

./cmake_build.py <path/to/installation>

and adding the following to the root CMakeLists.txt file of your project

list(APPEND CMAKE_PREFIX_PATH "<path/to/installation>")
find_package(wlplan)
...
target_link_libraries(<your_project> PRIVATE wlplan)

Graph Representations

The graph representations of planning tasks implemented thus far are

Name	WLPlan shorthand	Reference
Instance Learning Graph (ILG)	ilg	D. Chen, F. Trevizan, S. Thiébaux: Return to Tradition: Learning Reliable Heuristics with Classical Machine Learning. ICAPS 2024
Numeric Instance Learning Graph (νILG)	nilg	D. Chen, S. Thiébaux: Graph Learning for Numeric Planning. NeurIPS 2024

Feature Generators

The feature generators implemented thus far are

Name	WLPlan shorthand	Reference
Weisfeiler Leman (WL)/colour refinement	wl	D. Chen, F. Trevizan, S. Thiébaux: Return to Tradition: Learning Reliable Heuristics with Classical Machine Learning. ICAPS 2024
2-WL	kwl2	J. Cai, M. Fürer, N. Immerman: An optimal lower bound on the number of variables for graph identification. Combinatorica (1992)
Local 2-WL (2-LWL)	lwl2	C. Morris, K. Kersting, P. Mutzel: Glocalized Weisfeiler-Lehman Graph Kernels: Global-Local Feature Maps of Graphs. ICDM 2017
continuous-categorical WL (ccWL)	ccwl	D. Chen, S. Thiébaux: Graph Learning for Numeric Planning. NeurIPS 2024
individualised WL (iWL)	iwl	coming soon
normalised iWL (niWL)	niwl	coming soon

Usage

Examples for how to use the package include this self-contained notebook and test, as well as the GOOSE framework.