Sketch Learner

This repository contains code for learning policy sketches for a given set of training instances Q_Train from a class of problems Q over a common classical planning domain.

Environment

• Ubuntu 18.04
• Python 3.8.3

Installation

Step 1: Clone the repo

git clone git@github.com:drexlerd/sketch-learner.git

Step 2: Create Python3 virtual environment

python3 -m venv --prompt slearn .venv
source .venv/bin/activate
pip install -r requirements.txt

How to learn sketches for a new planning domain?

The following command prints the help message.

python3 learning/main.py --help

The following command, learns a sketch of width 1 for a set of planning problem over a common domain.

python3 learning/main.py --domain_filepath <path/to/pddl/domain> --problems_directory <path/to/pddl/problems> --workspace <path/to/workspace> --width 1

It is important that the planning problems are small. If you have a problem generator, then exhaustively generate small instances with the number of each object type in the range from 1 to 3.

Running the Learning Experiments (ICAPS2022)

Running the learning experiments requires setting the environment variable SKETCH_LEARNER_DIR to the location of the root of the repo. The experiments can be run in a slurm environment by running the following sequence of commands.

cd learning/experiments/scripts

./run_all_tractable.sh
./run_all_ipc2023.sh

Running the Testing Experiments (ICAPS2022)

Build the planner apptainers

cd testing/planners
./build-planners.sh

Run the experiments

cd testing/experiments
./experiment-siwr.py --all
./experiment-dual-bfws.py --all
./experiment-lama-first.py --all