GOOSE is a PDDL planning system that leverages graph learning. A high level description of the GOOSE framework here.
The main branch and recent releases currently do not support Graph Neural Networks (GNNs). This is because we showed in our ICAPS-24 paper that GNNs are inefficient and not effective for learning value functions for heuristic search. Nevertheless, GNNs can still be accessed from the following branches and the corresponding planning modes they support.
icaps24: lifted planningaaai24: domain-independent, ground and lifted planningneurips24: numeric planning
Instead, GOOSE uses Weisfeiler-Leman Features (WLFs) to construct embeddings of planning tasks for machine learning pipelines. The underlying code for doing so is located in the WLPlan library.
GOOSE can be installed and run with pretrained weights for the IPC23LT benchmarks in a matter of lines after cloning this repository.
# (1) Download the Apptainer image
apptainer pull goose.sif oras://ghcr.io/dillonzchen/goose:latest
# (2) Install benchmarks
git submodule update --init --recursive benchmarks
# (3) Plan with pretrained weights
./goose.sif plan benchmarks/ipc23lt/blocksworld/domain.pddl benchmarks/ipc23lt/blocksworld/testing/p0_30.pddl -m pretrained_models/ipc23lt-blocksworld.model
In step 3, blocksworld can be replaced with any other IPC23LT domain.
See further below for more information on how to train and plan with GOOSE for different settings.
First, install submodules
git submodule update --init --recursive
Then there are 3 possible ways to install GOOSE.
Download the image from the internet
apptainer pull goose.sif oras://ghcr.io/dillonzchen/goose:latest
Install Apptainer and then build the image
sudo apt-get install apptainer
python3 apptainer/build.py
You will first need the usual cpp packages
sudo apt-get install build-essential g++ cmake libboost-all-dev
[Optional] For numeric planning, you will also need Python2 as we use Numeric Fast Downward which requires it. To install from source (e.g. for Ubuntu 23.04 and above) and symlink to python2:
wget https://www.python.org/ftp/python/2.7.9/Python-2.7.9.tgz
sudo tar xzf Python-2.7.9.tgz
cd Python-2.7.9
sudo ./configure --enable-optimizations
sudo make altinstall
sudo ln -s /usr/local/bin/python2.7 /usr/local/bin/python2
Then create a virtual environment, activate it, install packages, and build cpp components. The setup has been tested with python versions 3.10 and higher.
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
./build
In case a virtual environment does not work, you can also try anaconda and specify a Python version:
conda create --name goose python=3.10.4
conda activate goose
pip install -r requirements.txt
./build
Call goose.sif train -h or python3 train.py -h for more detailed instructions. You will need the -s argument if you want to save the model.
- See below for recommended training configurations.
- To add your own datasets, follow the directory and file structure specified in
python3 train.py -h. - If you own a CPLEX license and want to train LP models faster, add it to PYTHONPATH and use the manual installation.
e.g.
./train.py benchmarks/neurips24/childsnack configurations/numeric.toml -s numeric_childsnack.model
Call goose.sif plan -h or python3 plan.py -h for more detailed instructions.
e.g.
./plan.py benchmarks/neurips24/childsnack/domain.pddl benchmarks/neurips24/childsnack/testing/p2_30.pddl -m numeric_childsnack.model
For classical planning, train with the configurations/classic.toml configuration file.
e.g. with Blocksworld
# Training
./train.py benchmarks/ipc23lt/blocksworld configurations/classic.toml -s blocksworld.model
# Planning
./plan.py benchmarks/ipc23lt/blocksworld/domain.pddl benchmarks/ipc23lt/blocksworld/testing/p1_01.pddl -m blocksworld.model
For lifted planning, train with the configurations/lifted.toml configuration file.
e.g. with Blocksworld
# Training
./train.py benchmarks/ipc23lt/blocksworld configurations/lifted.toml -s blocksworld-lifted.model
# Planning
./plan.py benchmarks/ipc23lt/blocksworld/domain.pddl benchmarks/ipc23lt/blocksworld/testing/p1_01.pddl -m blocksworld-lifted.model
For numeric planning, train with the configurations/numeric.toml configuration file.
e.g. with numeric Childsnack
# Training
./train.py benchmarks/neurips24/childsnack configurations/numeric.toml -s numeric_childsnack.model
# Planning
./plan.py benchmarks/neurips24/childsnack/domain.pddl benchmarks/neurips24/childsnack/testing/p2_30.pddl -m numeric_childsnack.model
- GNN: Graph Neural Network
- GOOSE: Graphs Optimised fOr Search Evaluation
- ICAPS: International Conference on Automated Planning and Scheduling
- IPC23LT: International Planning Competition 2023 Learning Track
- PDDL: Planning Domain Definition Language
- WLF: Weisfeiler-Leman Feature
If you use the code for your research paper, cite
@inproceedings{chen-trevizan-thiebaux-icaps2024,
author = {Dillon Z. Chen and
Felipe W. Trevizan and
Sylvie Thi{\'{e}}baux},
editor = {Sara Bernardini and
Christian Muise},
title = {Return to Tradition: Learning Reliable Heuristics with Classical Machine
Learning},
booktitle = {Proceedings of the Thirty-Fourth International Conference on Automated
Planning and Scheduling, {ICAPS} 2024, Banff, Alberta, Canada, June
1-6, 2024},
pages = {68--76},
publisher = {{AAAI} Press},
year = {2024},
}