AlphaNPI

Adapting the AlphaZero algorithm to remove the need of execution traces to train NPI.

Setup

You need to install the required Python packages.

cd alphanpi/

Then run the command:

pip install -r requirements.txt

Update the following environment variable:

export PYTHONPATH=$PWD:$PYTHONPATH

Training

cd trainings/

Run one of the scripts:

python train_recursive_sorting.py --tensorboard --verbose --save-model --save-results --save-model
python train_hanoi.py --tensorboard --verbose --save-model --save-results --save-model
python train_sorting_nohierarchy.py --tensorboard --verbose --save-model --save-results --save-model
python train_hanoi.py --tensorboard --verbose --save-model --save-results --save-model

For more information about the arguments that can be sent, see at trainings/README.md

Validation

The following allows to assert the results disclosed in the paper

cd validation/

Then run one of the scripts:

python validate_hanoi.py --verbose --save-results
python validate_recursive_sorting.py --verbose --save-results
python validate_sorting.py --verbose --save-results
python validate_sorting_nohierarchy.py --verbose --save-results

For more information about the arguments that can be sent, see at validation/README.md

Visualization

cd visualization/

To visualize a pre-trained model behavior for an environment env, run the script visualize_{env}.py. Set the load path at the beginning of the script to the path where the model of interest is saved.

Run one of the scripts:

python visualize_hanoi.py
python visualize_recursive_sorting.py
python visualize_sorting.py
python visualize_sorting_nohierarchy.py

Generate visualization

When a visualization script is executed, it generates a mcts.gv file under visualization/. The file contains a description of the tree in dot language.

If you don't already have graphviz installed, run command:

sudo apt-get install python3-pydot graphviz

To convert the .gv in .pdf file, use command:

dot -Tpdf mcts.gv -o mcts.pdf