This repository provides an implementation of our paper Grandmaster-Level Chess Without Search.
The recent breakthrough successes in machine learning are mainly attributed to scale: namely large-scale attention-based architectures and datasets of unprecedented scale. This paper investigates the impact of training at scale for chess. Unlike traditional chess engines that rely on complex heuristics, explicit search, or a combination of both, we train a 270M parameter transformer model with supervised learning on a dataset of 10 million chess games. We annotate each board in the dataset with action-values provided by the powerful Stockfish 16 engine, leading to roughly 15 billion data points. Our largest model reaches a Lichess blitz Elo of 2895 against humans, and successfully solves a series of challenging chess puzzles, without any domain-specific tweaks or explicit search algorithms. We also show that our model outperforms AlphaZero's policy and value networks (without MCTS) and GPT-3.5-turbo-instruct. A systematic investigation of model and dataset size shows that strong chess performance only arises at sufficient scale. To validate our results, we perform an extensive series of ablations of design choices and hyperparameters.
.
|
├── BayesElo - Elo computation (need to be installed)
|
├── checkpoints - Model checkpoints (need to be downloaded)
| ├── 136M
| ├── 270M
| └── 9M
|
├── data - Datasets (need to be downloaded)
| ├── eco_openings.csv
| ├── test
| ├── train
| └── puzzles.csv
|
├── lc0 - Leela Chess Zero (needs to be installed)
|
├── src
| ├── engines
| | ├── engine.py - Engine interface
| | ├── lc0_engine.py - Leela Chess Zero engine
| | ├── neural_engines.py - Neural engines
| | └── stockfish_engine.py - Stockfish engine
| |
| ├── bagz.py - Readers for our .bag data files
| ├── config.py - Experiment configurations
| ├── constants.py - Constants, interfaces, and types
| ├── data_loader.py - Data loader
| ├── metrics_evaluator.py - Metrics (e.g., Kendall's tau) evaluator
| ├── puzzles.py - Puzzle evaluation script
| ├── searchless_chess.ipynb - Model analysis notebook
| ├── tokenizer.py - Chess board tokenization
| ├── tournament.py - Elo tournament script
| ├── train.py - Example training + evaluation script
| ├── training.py - Training loop
| ├── training_utils.py - Training utility functions
| ├── transformer.py - Decoder-only Transformer
| └── utils.py - Utility functions
|
├── Stockfish - Stockfish (needs to be installed)
|
├── README.md
└── requirements.txt - Dependencies
Clone the source code into a local directory:
git clone https://github.com/google-deepmind/searchless_chess.git
cd searchless_chessThis repository requires Python 3.10.
pip install -r requirements.txt will install all required dependencies.
This is best done inside a conda environment.
To that end, install Anaconda.
Then, create and activate the conda environment:
conda create --name searchless_chess python=3.10
conda activate searchless_chessInstall pip and use it to install all the dependencies:
conda install pip
pip install -r requirements.txtIf you have a GPU available (highly recommended for fast training), then you can install JAX with CUDA support.
pip install --upgrade "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.htmlNote that the jax version must correspond to the existing CUDA installation you wish to use (CUDA 12 in the example above). Please see the JAX documentation for more details.
Download and compile the latest version of Stockfish (for Unix-like systems):
git clone https://github.com/official-stockfish/Stockfish.git
cd Stockfish/src
make -j profile-build ARCH=x86-64-avx2
cd ../..Follow the Lc0 download instructions, i.e.,
git clone -b release/0.30 --recurse-submodules https://github.com/LeelaChessZero/lc0.gitThen build the engine as described in the Lc0 build instructions.
We evaluate Lc0 with the largest-possible network from Lc0's model catalogue, i.e., the Large network.
To download that network, run the following command:
cd lc0/build/release
wget https://storage.lczero.org/files/768x15x24h-t82-swa-7464000.pb.gz
gzip -d 768x15x24h-t82-swa-7464000.pb.gz
cd ../../..To compute the Elos for the different agents, we require BayesElo, which can be installed as follows:
wget https://www.remi-coulom.fr/Bayesian-Elo/bayeselo.tar.bz2
tar -xvjf bayeselo.tar.bz2
cd BayesElo
make bayeselo
cd ..To download our datasets to the correct locations, run the following command:
cd data
./download.sh
cd ..We also provide the individual dataset download links in the following table (the action-value dataset is sharded into 2148 files due to its size and only the link to the first shard is listed below):
| Split | Action-Value | Behavioral Cloning | State-Value | Puzzles |
|---|---|---|---|---|
| Train | 1.2 GB (of 1.1 TB) | 34 GB | 36 GB | - |
| Test | 141 MB | 4.1 MB | 4.4 MB | 4.5 MB |
To download the pretrained models to the correct locations, run the following command:
cd checkpoints
./download.sh
cd ..Before running any code, make sure to activate the conda environment and set the PYTHONPATH:
conda activate searchless_chess
export PYTHONPATH=$(pwd)/..To train a model locally, run the following command:
cd src
python train.py
cd ..The model checkpoints will be saved to /checkpoints/local.
To evaluate a model's puzzle accuracy, run the following command:
cd src
python puzzles.py --num_puzzles 10 --agent=local
cd ..puzzles.py supports the following agents:
- the locally trained model:
local - the pretrained models:
9M,136M, and270M - the Stockfish engines:
stockfishandstockfish_all_moves - the Lc0 engines:
leela_chess_zero_depth_1,leela_chess_zero_policy_net, andleela_chess_zero_400_sims
To compute the Elo for the different agents, run the tournament to play games between them and then compute the Elo for the PGN file generated by the tournament (more information on BayesElo can be found here):
cd src
python tournament.py --num_games=200
cd ../BayesElo
./bayeselo
> ...
ResultSet>readpgn ../data/tournament.pgn
> N game(s) loaded, 0 game(s) with unknown result ignored.
ResultSet>elo
ResultSet-EloRating>mm
> 00:00:00,00
ResultSet-EloRating>exactdist
> 00:00:00,00
ResultSet-EloRating>ratings
> ...
cd ..To investigate the model's behavior (e.g., to compute the win percentage for all legal moves), start a notebook server and then open src/searchless_chess.ipynb in your browser:
jupyter notebook@article{ruoss2024grandmaster,
author = {Anian Ruoss and
Gr{\'{e}}goire Del{\'{e}}tang and
Sourabh Medapati and
Jordi Grau{-}Moya and
Li Kevin Wenliang and
Elliot Catt and
John Reid and
Tim Genewein},
title = {Grandmaster-Level Chess Without Search},
journal = {arXiv:2402.04494},
year = {2024}
}Copyright 2024 DeepMind Technologies Limited
All software is licensed under the Apache License, Version 2.0 (Apache 2.0); you may not use this file except in compliance with the Apache 2.0 license. You may obtain a copy of the Apache 2.0 license at: https://www.apache.org/licenses/LICENSE-2.0
The model weights are licensed under Creative Commons Attribution 4.0 (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode
Some portions of the dataset are in the public domain by a Creative Commons CC0 license from lichess.org. The remainder of the dataset is licensed under Creative Commons Attribution 4.0 (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode.
Unless required by applicable law or agreed to in writing, software and materials distributed under the Apache 2.0 or CC-BY licenses are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the licenses for the specific language governing permissions and limitations under those licenses.
This is not an official Google product.