Leela Chess Zero

Leela Chess Zero
Name	Leela Chess Zero
Author	Community project
Released	2018
Programming language	C++
Operating system	Linux, Windows, macOS
License	GPLv3

Leela Chess Zero

Leela Chess Zero is an open-source neural-network chess engine developed through a distributed research project involving contributors from communities associated with Stockfish, AlphaZero, DeepMind, Google DeepMind, FIDE, and academic groups at institutions such as MIT, Stanford University, University of California, Berkeley, and University of Cambridge. It implements self-play reinforcement learning inspired by work published by DeepMind and has been used in competitions involving players and engines connected to organizations like TCEC, Chess.com, FIDE World Championship, and tournaments hosted by ICGA, CCC and other events.

Overview

Leela Chess Zero is a neural-network-based chess engine emphasizing end-to-end learning similar to approaches pioneered in publications from DeepMind and teams at Google. The project integrates techniques from researchers affiliated with University of Toronto, Oxford University, ETH Zurich, Carnegie Mellon University, University of Montreal and others, while interfacing with chess software such as Arena (software), Cute Chess, Fritz (chess) and GUIs used by players like Magnus Carlsen, Hikaru Nakamura, Fabiano Caruana, Anish Giri and commentators at ChessBase. Its development drew attention from media outlets including The New York Times, BBC, Wired and journals like Nature and Science for comparisons with engines such as Stockfish, Komodo (chess) and commercial engines used in events by Grand Chess Tour.

Development and Training

The project's training pipeline used distributed computing volunteers coordinated via communities around GitHub, Discord, Reddit, Stack Overflow and lists associated with researchers at Google and laboratories at MIT CSAIL. Initial seeds referenced architectures publicized by DeepMind in papers presented at conferences like NeurIPS, ICML, ICLR and workshops at AAAI and IJCAI. Training datasets emerged from millions of self-play games validated against benchmarks maintained by FIDE rating lists and test suites used by teams at Stockfish development and academic groups from Princeton University, Yale University and University of Chicago. Contributors included hobbyists and professionals formerly associated with projects at IBM Research, Microsoft Research and institutions such as Harvard University and Caltech.

Architecture and Algorithms

Leela Chess Zero employs a convolutional residual network architecture influenced by designs used in papers from DeepMind authors and influenced researchers at Facebook AI Research, OpenAI, DeepMind AlphaGo teams, and academic groups at University of Oxford and EPFL. It uses Monte Carlo Tree Search methods similar to those applied in AlphaZero research, combined with loss functions and optimizers popularized in literature from Stanford AI Lab, Berkeley AI Research and labs at NYU and University of Toronto. The engine interfaces with the Universal Chess Interface used by GUIs such as WinBoard, XBoard, Shredder (chess) and incorporates evaluation heuristics compared in studies alongside techniques from Minimax algorithm research historically associated with work by researchers at Bell Labs and IBM. Contributors compared architectures with networks described in proceedings of NeurIPS and experiments replicated in preprints hosted by groups at Cornell University.

Competitive Performance and Matches

Leela Chess Zero has competed in markups of engine tournaments coordinated by organizations like TCEC, ICGA World Championship events and community matches on Chess.com and platforms used by grandmasters including Vladimir Kramnik, Viswanathan Anand, Ding Liren, Wesley So and commentators from ChessBase. Results were compared with engines such as Stockfish, Komodo, Houdini (chess) and commercial engines evaluated in matches reported by outlets like The Guardian and technical analyses presented at symposia organized by ACM and IEEE. Its match play influenced discussions at panels involving members of FIDE and organizers of events like the World Chess Championship and the Candidates Tournament.

Community and Development Model

The project uses an open development model on platforms such as GitHub, with source control workflows familiar to developers from Linux Kernel communities and contributors formerly associated with projects at Mozilla Foundation, Apache Software Foundation and companies like Google and Microsoft. Volunteers coordinate testing, network generation and resource donation through services and forums operated by Discord, Reddit, Chess.com, Lichess.org and event infrastructure providers used by TCEC. The contributor base includes researchers from universities such as ETH Zurich, University of Cambridge, Imperial College London, University of Oxford and independent developers whose prior work appeared in conferences like ICML, NeurIPS and ICLR.

Legal and Ethical Considerations

Legal and ethical discussions around the project referenced intellectual property debates involving organizations such as DeepMind, Google LLC, FIDE and contributors affiliated with academic institutions like Harvard, Stanford and MIT. Licensing under GPLv3 raised issues discussed in communities at GitHub and panels at conferences hosted by ACM and IEEE about reproducible research and open-source compliance. Ethical discourse involved stakeholders from FIDE, Chess.com, Lichess.org and academic ethicists from Oxford Internet Institute and Harvard Berkman Klein Center concerning fair play, engine-assisted cheating, and policies enforced in events organized by FIDE and major tournaments like the Grand Chess Tour.

Impact and Legacy

The project influenced ongoing research at institutions such as DeepMind, OpenAI, MIT, Stanford, University of Toronto and labs within Facebook AI Research and Microsoft Research, and affected how competitive organizations like TCEC, ICGA and platforms like Chess.com and Lichess.org approach engine development and anti-cheating policies. It inspired derivative projects and studies at conferences including NeurIPS, ICML and ICLR and fostered collaborations across communities in academia and industry such as teams from Google DeepMind and research groups at ETH Zurich.

Category:Chess engines