Off-Belief Learning

Introduction

This repo contains the implementation of the algorithm proposed in Off-Belief Learning, ICML 2021.

Update

[Mar 2022] We added the code for test time RL search, NeurIPS 2021 and test time Belief Fine-tuning, ICLR 2022 to this repo as they were built on the same policy and belief training infrastructure as off-belief learning. Check out pyhanabi/rl_search.py for the main entry point of the algorithm and searchcc/ for the backend code that implements search. Meanwhile, we included a simple single agent search baseline originally proposed in SPARTA, AAAI 2020, which can be accessed by running pyhanabi/sparta.py. Check pyhanabi/README for more detailed instructions.

[Feb 2022] We added new code in pyhanabi/bot to facilitate playing with the bot online. Checkout the pyhanabi/bot/README for more details.

[Feb 2022] We fixed a major pybind compatibility problem that has been preventing us from using newer pytorch version. Check the Environment Setup for more detail.

Environment Setup

We use conda/miniconda to manage environments.

conda create -n hanabi python=3.7
conda activate hanabi

pip install torch torchvision torchaudio

# install other dependencies
pip install psutil

# install a newer cmake if the current version is < 3.15
conda install -c conda-forge cmake

To help cmake find the proper libraries (e.g. libtorch), please either add the following lines to your .bashrc, or add it to a separate file and source it before you start working on the project.

# activate the conda environment
conda activate hanabi

# set path
CONDA_PREFIX=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
export CPATH=${CONDA_PREFIX}/include:${CPATH}
export LIBRARY_PATH=${CONDA_PREFIX}/lib:${LIBRARY_PATH}
export LD_LIBRARY_PATH=${CONDA_PREFIX}/lib:${LD_LIBRARY_PATH}

# avoid tensor operation using all cpu cores
export OMP_NUM_THREADS=1

The pybind here works with the pytorch1.10, the latest at the time of writing. If you use a newer version of pytorch that uses a different version of pybind, first check out the pybind module to use the corresponding version (the version can be found at pybind11 row here):

cd third_party/pybind11
git checkout $VERSION.XXX
cd ../..

Finally, to compile this repo:

# under project root
mkdir build
cd build
cmake ..
make -j10

Code Structure

For an overview of the training infrastructure, please refer to Figure 5 of the [Off-Belief Learning] (https://arxiv.org/pdf/2103.04000.pdf) paper.

hanabi-learning-environment is a modified version of the original HLE from Deepmind.

Notable modifications includes:

1. Card knowledge part of the observation encoding is changed to v0-belief, i.e. card knowledge normalized by the remaining public card count.

2. Functions to reset the game state with sampled hands.

rela (REinforcement Learning Assemly) is a set of tools for efficient batched neural network inference written in C++ with multi-threading.

rlcc implements the core of various algorithms. For example, the logic of fictitious transitions are implemented in r2d2_actor.cc. It also contains implementations of baselines such as other-play, VDN and IQL.

pyhanabi is the main entry point of the repo. It contains implementations for Q-network, recurrent DQN training, belief network and training, as well as some tools to analyze trained models.

Run the Code

Please refer to the README in pyhanabi for detailed instruction on how to train a model.

Download Models

To download the trained models used in the paper, go to models folder and run

sh download.sh

Due to agreement with BoardGameArena and Facebook policies, we are unable to release the "Clone Bot" models trained on the game data nor the datasets themselves.

Copyright

Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.

This source code is licensed under the license found in the LICENSE file in the root directory of this source tree.