This repository contains the code used for our MIT 6.883 Meta-Learning Fall 2020 course project.

It includes a python-only implementation of the ReBeL algorithm, along with our proposed ReBeL 2.0. The intention was to create a clean and easy-to-modify implementation for research purposes.

This code base was run and tested on Ubuntu 20.04. MacOS and Windows are not officially supported yet.

The code requires the following packages.

pytorch
jupyterlab
numpy
matplotlib
networkx
scipy
tqdm

To install:

1. Clone this repository to your local machine and.

 $ git clone https://github.com/damienwmartin/MetaLearning2020.git
 $ cd MetaLearning2020/

2. Follow the instructions to install pytorch listed on their website.

3. Install other dependencies using

pip:

 $ pip install -r requirements.txt

or conda:

 $ conda install -c conda-forge -c anaconda --file requirements.txt

The main implementation code is found in the Rebel/ folder.

The experiments/ folder contains jupyter notebooks for replicating the experiments in our project report.

The depreciated/ folder contains files that were used during the development process, and are no longer used. These files are currently staged for deletion.

More details about the implementation can be found here

The rebel() function defaults to vanilla RebeL unless you specify a node_lmit. For examples of usage, see Rebel/full_rebel.pyand experiments/tournaments.ipynb.

To use your own games, implement the functions in games/game_wrapper.py.

Original rebel paper can be found here and repository here