PPGS: Planning from Pixels in Environments with Combinatorially Hard Search Spaces

Environment Setup

• We recommend pipenv for creating and managing virtual environments (dependencies for other environment managers can be found in Pipfile)

git clone https://github.com/martius-lab/PPGS
cd ppgs
pipenv install
pipenv shell

• For simplicity, this codebase is ready for training on two of the three environments (IceSlider and DigitJump). They are part of the puzzlegen package, which we provide here, and can be simply installed with

pip install -e https://github.com/martius-lab/puzzlegen

• Offline datasets can be generated for training and validation. In the case of IceSlider we can use

python -m puzzlegen.extract_trajectories --record-dir /path/to/train_data --env-name ice_slider --start-level 0 --number-levels 1000 --max-steps 20 --n-repeat 20 --random 1
python -m puzzlegen.extract_trajectories --record-dir /path/to/test_data --env-name ice_slider --start-level 1000 --number-levels 1000 --max-steps 20 --n-repeat 5 --random 1

• Finally, we can add the paths to the extracted datasets in default_params.json as data_params.train_path and data_params.test_path. We should also set the name of the environment for validation in data_params.env_name ("ice_slider" for IceSlider or "digit_jump" for DigitJump).

• Training and evaluation are performed sequentially by running

python main.py

Configuration

All settings can be handled by editing default_config.json.