Jump to content

Leela Zero: Difference between revisions

From Wikipedia, the free encyclopedia
Browse history interactively
← Previous edit
Content deleted Content added
VisualWikitext
Citation bot (talk | contribs)
Bots
5,391,599 edits
Altered template type. Add: class, eprint, authors 1-1. Removed URL that duplicated identifier. Removed access-date with no URL. Removed parameters. Some additions/deletions were parameter name changes. | Use this bot. Report bugs. | Suggested by Headbomb | #UCB_toolbar
 
(6 intermediate revisions by 3 users not shown)
Line 15: Line 15:
| developer = Gian-Carlo Pascutto
| developer = Gian-Carlo Pascutto
| released = {{Start date and age|2017|10|25|df=yes}}
| released = {{Start date and age|2017|10|25|df=yes}}
| discontinued =
| discontinued = yes
| latest release version = 0.17
| latest release version = 0.17
| latest release date ={{Start date and age|2019|04|04|df=yes}}
| latest release date = {{Start date and age|2019|04|04|df=yes}}
| latest preview version =
| latest preview version =
| latest preview date = <!-- {{Start date and age|YYYY|MM|DD|df=yes/no}} -->
| latest preview date = <!-- {{Start date and age|YYYY|MM|DD|df=yes/no}} -->
| programming language =
| programming language = [[C++]]
| operating system =
| operating system =
| platform =
| platform =
Line 34: Line 34:
| AsOf =
| AsOf =
}}
}}
'''Leela Zero''' is a [[Free and open-source software|free and open-source]] [[computer Go]] program released on 25 October 2017. It is developed by Belgian programmer [[Gian-Carlo Pascutto]],<ref>{{cite web|url=http://www.xinhuanet.com/english/2018-04/09/c_137097436.htm|archive-url=https://web.archive.org/web/20180409062437/http://www.xinhuanet.com/english/2018-04/09/c_137097436.htm|url-status=dead|archive-date=April 9, 2018|title=Feature: One man's Go program looks to remake AlphaGo Zero - and beyond|publisher=[[Xinhuanet]]|date=9 April 2018|access-date=28 April 2018}}</ref><ref>{{cite web|url=http://sports.xinhuanet.com/c/2018-02/05/c_1122370241.htm|title=围棋AI"丽拉"获赞接近职业棋手水准，它的作者竟是一个不太会下棋的程序员|publisher=[[Xinhuanet]]|date=5 February 2018|access-date=27 April 2018|language=zh|archive-date=11 February 2018|archive-url=https://web.archive.org/web/20180211113835/http://sports.xinhuanet.com/c/2018-02/05/c_1122370241.htm|url-status=dead}}</ref><ref name="Xinhuanet20180408">{{cite web|url=http://www.xinhuanet.com/sports/2018-04/08/c_1122648768.htm|archive-url=https://web.archive.org/web/20180409051235/http://www.xinhuanet.com/sports/2018-04/08/c_1122648768.htm|url-status=dead|archive-date=April 9, 2018|title=更开放，更共享，比利时围棋AI"丽拉·元"重塑"阿尔法元"|publisher=[[Xinhuanet]]|date=8 April 2018|access-date=27 April 2018|language=zh}}</ref> the author of chess engine [[Sjeng (software)|Sjeng]] and Go engine [[Leela (software)|Leela]].<ref name="baduknews">{{cite web|url=http://baduknews.com/news/view.php?idx=236|title=프로 수준급 인공지능 바둑 프로그램 '릴라(Leela)' 무료 공개|publisher=Baduk News|language=Korean|date=23 February 2017|access-date=27 April 2018|archive-url=https://web.archive.org/web/20180706021721/http://baduknews.com/news/view.php?idx=236|archive-date=6 July 2018|url-status=dead}}</ref><ref>{{cite web|url=https://www.cyberoro.com/board/oro_view.oro?bd_div=1&bd_num=16436|title=릴라의 출현과 온라인 대국의 비극적인 종말...|publisher=Cyberoro|date=3 March 2017|access-date=27 April 2018|language=Korean}}</ref>
'''Leela Zero''' is a [[Free and open-source software|free and open-source]] [[computer Go]] program released on 25 October 2017. It is developed by Belgian programmer [[Gian-Carlo Pascutto]],<ref>{{cite web|url=http://www.xinhuanet.com/english/2018-04/09/c_137097436.htm|archive-url=https://web.archive.org/web/20180409062437/http://www.xinhuanet.com/english/2018-04/09/c_137097436.htm|url-status=dead|archive-date=April 9, 2018|title=Feature: One man's Go program looks to remake AlphaGo Zero - and beyond|publisher=[[Xinhuanet]]|date=9 April 2018|access-date=28 April 2018}}</ref><ref>{{cite web|url=http://sports.xinhuanet.com/c/2018-02/05/c_1122370241.htm|title=围棋AI"丽拉"获赞接近职业棋手水准，它的作者竟是一个不太会下棋的程序员|publisher=[[Xinhuanet]]|date=5 February 2018|access-date=27 April 2018|language=zh|archive-date=11 February 2018|archive-url=https://web.archive.org/web/20180211113835/http://sports.xinhuanet.com/c/2018-02/05/c_1122370241.htm|url-status=dead}}</ref><ref name="Xinhuanet20180408">{{cite web|url=http://www.xinhuanet.com/sports/2018-04/08/c_1122648768.htm|archive-url=https://web.archive.org/web/20180409051235/http://www.xinhuanet.com/sports/2018-04/08/c_1122648768.htm|url-status=dead|archive-date=April 9, 2018|title=更开放，更共享，比利时围棋AI"丽拉·元"重塑"阿尔法元"|publisher=[[Xinhuanet]]|date=8 April 2018|access-date=27 April 2018|language=zh}}</ref> the author of chess engine [[Sjeng (software)|Sjeng]] and Go engine [[Leela (software)|Leela]].<ref name="baduknews">{{cite web|url=http://baduknews.com/news/view.php?idx=236|title=프로 수준급 인공지능 바둑 프로그램 '릴라(Leela)' 무료 공개|publisher=Baduk News|language=Korean|date=23 February 2017|access-date=27 April 2018|archive-url=https://web.archive.org/web/20180706021721/http://baduknews.com/news/view.php?idx=236|archive-date=6 July 2018|url-status=dead}}</ref><ref>{{cite web|url=https://www.cyberoro.com/board/oro_view.oro?bd_div=1&bd_num=16436|title=릴라의 출현과 온라인 대국의 비극적인 종말...|publisher=Cyberoro|date=3 March 2017|access-date=27 April 2018|language=Korean}}</ref>


Leela Zero's algorithm is based on [[DeepMind]]'s 2017 paper about [[AlphaGo Zero]].<ref name="Xinhuanet20180408"/><ref>{{cite web|url=https://github.com/gcp/leela-zero|title=leela-zero|publisher=[[GitHub]]|date=|access-date=27 April 2018}}</ref>
Leela Zero's algorithm is based on [[DeepMind]]'s 2017 paper about [[AlphaGo Zero]].<ref name="Xinhuanet20180408"/><ref name=":0">{{cite web|url=https://github.com/gcp/leela-zero|title=leela-zero|publisher=[[GitHub]]|date=|access-date=27 April 2018}}</ref>
Unlike the original Leela, which has a lot of human knowledge and heuristics programmed into it, the program code in Leela Zero only knows the basic rules and nothing more. The knowledge that makes Leela Zero a strong player is contained in a [[artificial neural network|neural network]], which is trained based on the results of previous games that the program played.<ref name="egf interview">{{cite web|url=https://www.eurogofed.org/?id=205| title=Gian-Carlo Pascutto - The man behind LeelaZero |publisher=[[European Go Federation]] |date=24 May 2018 |access-date=27 May 2018}}</ref>
Unlike the original Leela, which has a lot of human knowledge and heuristics programmed into it, the program code in Leela Zero only knows the basic rules and nothing more. The knowledge that makes Leela Zero a strong player is contained in a [[artificial neural network|neural network]], which is trained based on the results of previous games that the program played.<ref name="egf interview">{{cite web|url=https://www.eurogofed.org/?id=205| title=Gian-Carlo Pascutto - The man behind LeelaZero |publisher=[[European Go Federation]] |date=24 May 2018 |access-date=27 May 2018}}</ref>


Leela Zero is trained by a distributed effort, which is coordinated at the Leela Zero website. Members of the community provide computing resources by running the client, which generates [[Self-play (reinforcement learning technique)|self-play]] games and submits them to the server. The self-play games are used to train newer networks. Generally, over 500 clients have connected to the server to contribute resources.<ref name="egf interview"/> The community has provided high quality code contributions as well.<ref name="egf interview"/>
Leela Zero is trained by a distributed effort, which is coordinated at the Leela Zero website. Members of the community provide computing resources by running the client, which generates [[Self-play (reinforcement learning technique)|self-play]] games and submits them to the server. The self-play games are used to train newer networks. Generally, over 500 clients have connected to the server to contribute resources.<ref name="egf interview"/> The community has provided high quality code contributions as well.<ref name="egf interview"/>


== Version history ==
Leela Zero finished third at the BerryGenomics Cup World AI Go Tournament in [[Fuzhou]], [[Fujian]], China on 28 April 2018.<ref>{{cite web|url=http://sports.sina.com.cn/go/2018-04-28/doc-ifzvpatq5885312.shtml |title=世界AI大赛决赛腾讯内战 凤凰2-1绝艺夺冠|publisher=[[sina.com.cn]]|date=28 April 2018|access-date=28 April 2018|language=zh}} </ref> ''[[The New Yorker]]'' at the end of 2018 characterized Leela and Leela Zero as "the world’s most successful open-source Go engines".<ref>{{cite magazine |title=How the Artificial-Intelligence Program AlphaZero Mastered Its Games |url=https://www.newyorker.com/science/elements/how-the-artificial-intelligence-program-alphazero-mastered-its-games |access-date=31 December 2018 |magazine=[[The New Yorker]] |date=2018}}</ref>
Leela Zero finished third at the BerryGenomics Cup World AI Go Tournament in [[Fuzhou]], [[Fujian]], China on 28 April 2018.<ref>{{cite web |date=28 April 2018 |title=世界AI大赛决赛腾讯内战 凤凰2-1绝艺夺冠 |url=http://sports.sina.com.cn/go/2018-04-28/doc-ifzvpatq5885312.shtml |access-date=28 April 2018 |publisher=[[sina.com.cn]] |language=zh}}</ref> ''[[The New Yorker]]'' at the end of 2018 characterized Leela and Leela Zero as "the world’s most successful open-source Go engines".<ref>{{cite magazine |date=2018 |title=How the Artificial-Intelligence Program AlphaZero Mastered Its Games |url=https://www.newyorker.com/science/elements/how-the-artificial-intelligence-program-alphazero-mastered-its-games |access-date=31 December 2018 |magazine=[[The New Yorker]]}}</ref>


In early 2018, another team branched [[Leela Chess Zero]] from the same code base, also to verify the methods in the [[AlphaZero]] paper as applied to the game of chess. AlphaZero's use of Google TPUs was replaced by a crowd-sourcing infrastructure and the ability to use graphics card GPUs via the OpenCL library. Even so, it is expected to take a year of crowd-sourced training to make up for the dozen hours that AlphaZero was allowed to train for its chess match in the paper.<ref name="SilverApril2018">{{cite web |last1=Silver |first1=Albert |title=Leela Chess Zero: AlphaZero for the PC |url=https://en.chessbase.com/post/leela-chess-zero-alphazero-for-the-pc |website=Chess News |access-date=11 June 2018 |date=26 April 2018}}</ref>
In early 2018, another team branched [[Leela Chess Zero]] from the same code base, also to verify the methods in the [[AlphaZero]] paper as applied to the game of chess. AlphaZero's use of Google TPUs was replaced by a crowd-sourcing infrastructure and the ability to use graphics card GPUs via the OpenCL library. Even so, it is expected to take a year of crowd-sourced training to make up for the dozen hours that AlphaZero was allowed to train for its chess match in the paper.<ref name="SilverApril2018">{{cite web |last1=Silver |first1=Albert |date=26 April 2018 |title=Leela Chess Zero: AlphaZero for the PC |url=https://en.chessbase.com/post/leela-chess-zero-alphazero-for-the-pc |access-date=11 June 2018 |website=Chess News}}</ref>

The distributed training server was shut down on 2021-02-15, marking the end of Leela Zero project. The page now directs visitors to [[KataGo]] and SAI.<ref>{{cite arXiv |last1=Morandin |first1=Francesco |title=SAI: a Sensible Artificial Intelligence that plays with handicap and targets high scores in 9x9 Go (extended version) |date=2019-11-26 |eprint=1905.10863 |last2=Amato |first2=Gianluca |last3=Fantozzi |first3=Marco |last4=Gini |first4=Rosa |last5=Metta |first5=Carlo |last6=Parton |first6=Maurizio|class=cs.AI }}</ref>

The model sizes increased steadily over time. The first released model has hash name d645af97, size 1x8 (1 layer, 8 channels), and released at 2017-11-10 13:04. The last released model has hash name 0e9ea880, size 40x256, and was released at 2021-02-15 09:04.<ref>{{Cite web |title=Leela Zero |url=https://zero.sjeng.org/ |access-date=2024-09-21 |website=zero.sjeng.org}}</ref>

== Technology ==
Leela Zero is an (almost) exact replication of [[AlphaGo Zero]] in both training process and architecture.<ref name="Nature2017">{{cite journal |last1=Silver |first1=David |author-link1=David Silver (programmer) |last2=Schrittwieser |first2=Julian |last3=Simonyan |first3=Karen |last4=Antonoglou |first4=Ioannis |last5=Huang |first5=Aja |author-link5=Aja Huang |last6=Guez |first6=Arthur |last7=Hubert |first7=Thomas |last8=Baker |first8=Lucas |last9=Lai |first9=Matthew |last10=Bolton |first10=Adrian |last11=Chen |first11=Yutian |author-link11=Chen Yutian |last12=Lillicrap |first12=Timothy |last13=Fan |first13=Hui |author-link13=Fan Hui |last14=Sifre |first14=Laurent |last15=Driessche |first15=George van den |date=19 October 2017 |title=Mastering the game of Go without human knowledge |url=http://discovery.ucl.ac.uk/10045895/1/agz_unformatted_nature.pdf |url-status=live |journal=[[Nature (journal)|Nature]] |volume=550 |issue=7676 |pages=354–359 |bibcode=2017Natur.550..354S |doi=10.1038/nature24270 |issn=0028-0836 |pmid=29052630 |s2cid=205261034 |archive-url=https://web.archive.org/web/20180718225914/http://discovery.ucl.ac.uk/10045895/1/agz_unformatted_nature.pdf |archive-date=18 July 2018 |access-date=2 September 2019 |first16=Thore |last16=Graepel |first17=Demis |last17=Hassabis |author-link17=Demis Hassabis}}{{closed access}}</ref>

The training process is [[Monte Carlo tree search|Monte-Carlo Tree Search]] with [[self-play]], exactly the same as AlphaGo Zero.

The architecture is the same as AlphaGo Zero (with one difference). Consider the last released model, 0e9ea880. It has 47 million parameters, and the following architecture:<ref>{{Cite web |last=Du |first=Haoxing |date=2023-03-01 |title=Inside the mind of a superhuman Go model: How does Leela Zero read ladders? |url=https://www.lesswrong.com/posts/FF8i6SLfKb4g7C4EL/inside-the-mind-of-a-superhuman-go-model-how-does-leela-zero-2 |language=en}}</ref><ref name=":0" />

* The stem of the network takes as input a 18x19x19 tensor representation of the Go board.
** 8 channels are the positions of the current player's stones from the last eight time steps. (1 if there is a stone, 0 otherwise. If the time step go before the beginning of the game, then 0 in all positions.)
** 8 channels are the positions of the other player's stones from the last eight time steps.
** 1 channel is all 1 if black is to move, and 0 otherwise.
** 1 channel is all 1 if white is to move, and 0 otherwise. (This channel is not present in the original AlphaGo Zero)
* The body is a [[Residual neural network|ResNet]] with 40 residual blocks and 256 channels.
* There are two heads, a policy head and a value head.
** Policy head outputs a [[Multinomial logistic regression|logit]] array of size <math>19 \times 19 + 1</math>, representing the logit of making a move in one of the points, plus the logit of [[Rules of Go#Moving|passing]].
** Value head outputs a number in the range <math>(-1, +1)</math>, representing the expected score for the current player. -1 represents current player losing, and +1 winning.


==References==
==References==
Line 49: Line 71:


==External links==
==External links==
*{{Official|http://zero.sjeng.org}}
*{{Official website|http://zero.sjeng.org}}
*[https://github.com/gcp/leela-zero Leela Zero] on [[GitHub]]
*[https://github.com/gcp/leela-zero Leela Zero] on [[GitHub]]
*[https://senseis.xmp.net/?LeelaZero Leela Zero] on [[Sensei's Library]]
*[https://senseis.xmp.net/?LeelaZero Leela Zero] on [[Sensei's Library]]
Line 55: Line 77:


{{Go (game)}}
{{Go (game)}}

[[Category:Go engines]]
[[Category:Go engines]]
[[Category:Free and open-source software]]
[[Category:Free and open-source software]]
Line 60: Line 83:
[[Category:Applied machine learning]]
[[Category:Applied machine learning]]
[[Category:Software using the GPL license]]
[[Category:Software using the GPL license]]
[[Category:Free software programmed in C++]]

Latest revision as of 07:48, 25 September 2024

Original author(s)Gian-Carlo Pascutto
Developer(s)Gian-Carlo Pascutto
Initial release25 October 2017; 7 years ago (2017-10-25)
Final release
0.17 / 4 April 2019; 5 years ago (2019-04-04)
Repository
Written inC++
TypeGo software
LicenseGPL-3.0
Websitezero.sjeng.org

Leela Zero is a free and open-source computer Go program released on 25 October 2017. It is developed by Belgian programmer Gian-Carlo Pascutto,[1][2][3] the author of chess engine Sjeng and Go engine Leela.[4][5]

Leela Zero's algorithm is based on DeepMind's 2017 paper about AlphaGo Zero.[3][6] Unlike the original Leela, which has a lot of human knowledge and heuristics programmed into it, the program code in Leela Zero only knows the basic rules and nothing more. The knowledge that makes Leela Zero a strong player is contained in a neural network, which is trained based on the results of previous games that the program played.[7]

Leela Zero is trained by a distributed effort, which is coordinated at the Leela Zero website. Members of the community provide computing resources by running the client, which generates self-play games and submits them to the server. The self-play games are used to train newer networks. Generally, over 500 clients have connected to the server to contribute resources.[7] The community has provided high quality code contributions as well.[7]

Version history

[edit]

Leela Zero finished third at the BerryGenomics Cup World AI Go Tournament in Fuzhou, Fujian, China on 28 April 2018.[8] The New Yorker at the end of 2018 characterized Leela and Leela Zero as "the world’s most successful open-source Go engines".[9]

In early 2018, another team branched Leela Chess Zero from the same code base, also to verify the methods in the AlphaZero paper as applied to the game of chess. AlphaZero's use of Google TPUs was replaced by a crowd-sourcing infrastructure and the ability to use graphics card GPUs via the OpenCL library. Even so, it is expected to take a year of crowd-sourced training to make up for the dozen hours that AlphaZero was allowed to train for its chess match in the paper.[10]

The distributed training server was shut down on 2021-02-15, marking the end of Leela Zero project. The page now directs visitors to KataGo and SAI.[11]

The model sizes increased steadily over time. The first released model has hash name d645af97, size 1x8 (1 layer, 8 channels), and released at 2017-11-10 13:04. The last released model has hash name 0e9ea880, size 40x256, and was released at 2021-02-15 09:04.[12]

Technology

[edit]

Leela Zero is an (almost) exact replication of AlphaGo Zero in both training process and architecture.[13]

The training process is Monte-Carlo Tree Search with self-play, exactly the same as AlphaGo Zero.

The architecture is the same as AlphaGo Zero (with one difference). Consider the last released model, 0e9ea880. It has 47 million parameters, and the following architecture:[14][6]

  • The stem of the network takes as input a 18x19x19 tensor representation of the Go board.
    • 8 channels are the positions of the current player's stones from the last eight time steps. (1 if there is a stone, 0 otherwise. If the time step go before the beginning of the game, then 0 in all positions.)
    • 8 channels are the positions of the other player's stones from the last eight time steps.
    • 1 channel is all 1 if black is to move, and 0 otherwise.
    • 1 channel is all 1 if white is to move, and 0 otherwise. (This channel is not present in the original AlphaGo Zero)
  • The body is a ResNet with 40 residual blocks and 256 channels.
  • There are two heads, a policy head and a value head.
    • Policy head outputs a logit array of size , representing the logit of making a move in one of the points, plus the logit of passing.
    • Value head outputs a number in the range , representing the expected score for the current player. -1 represents current player losing, and +1 winning.

References

[edit]
  1. ^ "Feature: One man's Go program looks to remake AlphaGo Zero - and beyond". Xinhuanet. 9 April 2018. Archived from the original on April 9, 2018. Retrieved 28 April 2018.
  2. ^ "围棋AI"丽拉"获赞接近职业棋手水准，它的作者竟是一个不太会下棋的程序员" (in Chinese). Xinhuanet. 5 February 2018. Archived from the original on 11 February 2018. Retrieved 27 April 2018.
  3. ^ a b "更开放，更共享，比利时围棋AI"丽拉·元"重塑"阿尔法元"" (in Chinese). Xinhuanet. 8 April 2018. Archived from the original on April 9, 2018. Retrieved 27 April 2018.
  4. ^ "프로 수준급 인공지능 바둑 프로그램 '릴라(Leela)' 무료 공개" (in Korean). Baduk News. 23 February 2017. Archived from the original on 6 July 2018. Retrieved 27 April 2018.
  5. ^ "릴라의 출현과 온라인 대국의 비극적인 종말..." (in Korean). Cyberoro. 3 March 2017. Retrieved 27 April 2018.
  6. ^ a b "leela-zero". GitHub. Retrieved 27 April 2018.
  7. ^ a b c "Gian-Carlo Pascutto - The man behind LeelaZero". European Go Federation. 24 May 2018. Retrieved 27 May 2018.
  8. ^ "世界AI大赛决赛腾讯内战 凤凰2-1绝艺夺冠" (in Chinese). sina.com.cn. 28 April 2018. Retrieved 28 April 2018.
  9. ^ "How the Artificial-Intelligence Program AlphaZero Mastered Its Games". The New Yorker. 2018. Retrieved 31 December 2018.
  10. ^ Silver, Albert (26 April 2018). "Leela Chess Zero: AlphaZero for the PC". Chess News. Retrieved 11 June 2018.
  11. ^ Morandin, Francesco; Amato, Gianluca; Fantozzi, Marco; Gini, Rosa; Metta, Carlo; Parton, Maurizio (2019-11-26). "SAI: a Sensible Artificial Intelligence that plays with handicap and targets high scores in 9x9 Go (extended version)". arXiv:1905.10863 [cs.AI].
  12. ^ "Leela Zero". zero.sjeng.org. Retrieved 2024-09-21.
  13. ^ Silver, David; Schrittwieser, Julian; Simonyan, Karen; Antonoglou, Ioannis; Huang, Aja; Guez, Arthur; Hubert, Thomas; Baker, Lucas; Lai, Matthew; Bolton, Adrian; Chen, Yutian; Lillicrap, Timothy; Fan, Hui; Sifre, Laurent; Driessche, George van den; Graepel, Thore; Hassabis, Demis (19 October 2017). "Mastering the game of Go without human knowledge" (PDF). Nature. 550 (7676): 354–359. Bibcode:2017Natur.550..354S. doi:10.1038/nature24270. ISSN 0028-0836. PMID 29052630. S2CID 205261034. Archived (PDF) from the original on 18 July 2018. Retrieved 2 September 2019.Closed access icon
  14. ^ Du, Haoxing (2023-03-01). "Inside the mind of a superhuman Go model: How does Leela Zero read ladders?".
[edit]
Retrieved from "https://en.wikipedia.org/enwiki/w/index.php?title=Leela_Zero&oldid=1247651283"