imported>InternetArchiveBot：补救2个来源，并将0个来源标记为失效。) #IABot (v2.0.9.5

2025-02-06T17:02:22Z

补救2个来源，并将0个来源标记为失效。) #IABot (v2.0.9.5

新页面

{{NoteTA
|G1 = IT
}}
{{Infobox software
| name = ELF OpenGo
| title = ELF OpenGo
| logo = 
| logo caption =
| logo_size =
| logo_alt =
| screenshot = 
| caption =
| screenshot_size =
| screenshot_alt =
| collapsible =
| author = {{Plainlist|
* 田渊栋（Yuandong Tian）
* Jerry Ma
* Qucheng Gong
* Shubho Sengupta
* 陈卓远（Zhuoyuan Chen）
* C. Lawrence Zitnick<ref name="elf-github">{{Cite web |url=https://github.com/pytorch/elf |title=pytorch/ELF: ELF: a platform for game research |accessdate=2018-05-07 |archive-date=2018-05-03 |archive-url=https://web.archive.org/web/20180503010229/https://github.com/pytorch/elf |dead-url=no }}</ref>
}}
| developer =
| released = v0（{{Start date and age|2018|05|02|df=yes}}）
| discontinued =
| programming language = [[C++]]、[[Python]]、[[C语言]]
| operating system =
| platform =
| size =
| language =
| language count = 
| language footnote =
| genre = [[電腦圍棋]]
| license = [[BSD许可证]]
| repo = {{URL|https://github.com/pytorch/elf}}
| website = {{URL|https://facebook.ai/developers/tools/elf}}
| standard =
| AsOf =
}}
'''ELF OpenGo'''是[[Facebook]] AI Research團隊（{{lang|en|FAIR}}）所開發的[[電腦圍棋]]軟體及所釋出的資料<ref name="announcement">{{Cite web |url=https://research.fb.com/facebook-open-sources-elf-opengo/ |title=Facebook Open Sources ELF OpenGo |accessdate=2018-05-08 |date=2018-05-02 |archive-date=2018-05-07 |archive-url=https://web.archive.org/web/20180507163202/https://research.fb.com/facebook-open-sources-elf-opengo/ |dead-url=no }}</ref>。

== 簡介 ==
{{See also|AlphaGo Zero}}
ELF OpenGo是[[Facebook]] AI Research團隊（{{lang|en|FAIR}}）依照[[Google DeepMind]]在科學期刊《[[自然 (期刊)|自然]]》上對於[[AlphaGo Zero]]所發表的論文《{{lang|en|Mastering the game of Go without human knowledge}}<ref name="agz-nature">{{Cite journal |last=Silver |first=David |last2=Schrittwieser |first2=Julian |last3=Simonyan |first3=Karen |last4=Antonoglou |first4=Ioannis |last5=Huang |first5=Aja |last6=Guez |first6=Arthur |last7=Hubert |first7=Thomas |last8=Baker |first8=Lucas |last9=Lai |first9=Matthew |last10=Bolton |first10=Adrian |last11=Chen |first11=Yutian |title=Mastering the game of Go without human knowledge |url=https://www.nature.com/articles/nature24270 |journal=Nature |language=en |date=2017-10 |volume=550 |issue=7676 |doi=10.1038/nature24270 |issn=1476-4687 |access-date=2018-05-06 |archive-date=2018-04-12 |archive-url=https://web.archive.org/web/20180412102056/https://www.nature.com/articles/nature24270 |dead-url=no }}</ref>》與[[AlphaZero]]的論文《{{lang|en|Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm}}<ref name="az-arxiv">{{Citation|title=Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm|url=https://arxiv.org/abs/1712.01815|date=2017-12-05|accessdate=2025-02-04|doi=10.48550/arXiv.1712.01815|first=David|last=Silver|first2=Thomas|last2=Hubert|first3=Julian|last3=Schrittwieser|first4=Ioannis|last4=Antonoglou|first5=Matthew|last5=Lai|first6=Arthur|last6=Guez|first7=Marc|last7=Lanctot|first8=Laurent|last8=Sifre|first9=Dharshan|last9=Kumaran|archive-date=2017-12-08|archive-url=https://web.archive.org/web/20171208234501/https://arxiv.org/abs/1712.01815|dead-url=no}}</ref>》所實做出的[[開源]]電腦圍棋程式<ref name="elf">{{Cite web |url=https://facebook.ai/developers/tools/elf |title=ELF | Game Research Platform | Facebook AI |accessdate=2018-05-06 |archive-date=2018-05-06 |archive-url=https://web.archive.org/web/20180506142732/https://facebook.ai/developers/tools/elf |dead-url=yes }}</ref>，也就是不使用人類棋譜與累積的圍棋知識，僅實做[[圍棋規則]]，使用單一[[類神經網路]]從自我對弈中學習（不像[[AlphaGo]]以人類角度思考，設計了Policy Network與Value Network）。

訓練網路使用20 blocks x 224 filters，在2000個[[GPU]]下訓練兩週<ref name="elf"/>，相較AlphaGo Zero使用的20 blocks x 256 filters版本略小一些（AlphaGo Zero另外還有40 blocks x 256 filters）。

由於[[Facebook]]所擁有的計算資源，產生出高品質的訓練網路資料與對局棋譜<ref name="release">{{Cite web |url=https://github.com/pytorch/ELF/releases |title=Releases · pytorch/ELF |accessdate=2018-05-06 |archive-date=2019-02-15 |archive-url=https://web.archive.org/web/20190215234740/https://github.com/pytorch/ELF/releases |dead-url=no }}</ref>{{notetag|包含所有電腦對弈棋譜，以及12盤人類棋手願意公開的棋譜資訊（14盤中有2盤因尊重棋手意願不公開<ref name="release"/>）。另外在棋譜中未納入人類棋手的名字<ref name="release"/>。}}，許多基於相同演算法或是AlphaGo相關論文內容的圍棋軟體都積極測試ELF OpenGo所提供的訓練資料<ref name="leelazero-release">{{Cite web |url=https://github.com/gcp/leela-zero/releases |title=Releases · gcp/leela-zero |accessdate=2018-05-09 |archive-date=2019-10-16 |archive-url=https://web.archive.org/web/20191016104855/https://github.com/leela-zero/leela-zero/releases |dead-url=no }}</ref><ref name="ray-release">{{Cite web |url=https://github.com/zakki/Ray/releases |title=Releases · zakki/Ray |accessdate=2018-05-07 |archive-date=2019-02-15 |archive-url=https://web.archive.org/web/20190215234740/https://github.com/zakki/Ray/releases |dead-url=no }}</ref>。

== 成績 ==
=== 對電腦 ===
==== Leela Zero ====
[[Leela Zero]]是目前少數有公開程式碼並公開訓練網路資料的[[圍棋軟體]]，且仍然有志願者持續投入資源計算演化，故經常被當作其他圍棋軟體的基準。

由[[Facebook]]自行測試，ELF OpenGo與Leela Zero對戰的成績為198:2<ref name="announcement"/>{{notetag|Leela Zero的訓練網路為2018年4月25日所訓練出的第128代，Hash值為<code>158603eb</code><ref name="leelazero-project">{{Cite web |url=https://zero.sjeng.org/ |title=Leela Zero |accessdate=2018-05-06 |archive-date=2018-05-05 |archive-url=https://web.archive.org/web/20180505160607/http://zero.sjeng.org/ |dead-url=no }}</ref>的版本。}}。在Leela Zero的進度網站上也經常會比較現有訓練網路與ELF OpenGo的比較<ref name="leelazero-project"/>。

==== CGOS ====
志願者在[[CGOS]]上使用[[Leela Zero]]的程式引擎以及由ELF OpenGo公開的訓練網路（v0）轉換成Leela Zero格式的訓練網路（即Hash值<code>62b5417b</code>的訓練網路，帳號<code>LZ_62b541_ELF_1600</code>）進行對弈測試，但由於目前在CGOS上測試的不是ELF Go的程式引擎，不清楚ELF OpenGo的訓練資料在轉換後的影響。{{As of|2018|5|30}}，對弈已經超過1000盤，BayesElo分數約3770分<ref name="bayes-19x19">{{Cite web |url=http://www.yss-aya.com/cgos/19x19/bayes.html |title=19x19 All Time Ranks |accessdate=2018-05-30 |archive-date=2018-01-17 |archive-url=https://web.archive.org/web/20180117032048/http://www.yss-aya.com/cgos/19x19/bayes.html |dead-url=no }}</ref>。

=== 對人 ===
[[Facebook]]與[[韩国棋院]]合作，以[[圍棋規則比較|中國規則]]與韓國的世界頂尖棋手對弈（貼目7.5目），在電腦每步限制50秒（使用單機單張[[NVIDIA Tesla|NVIDIA Tesla V100]]），人類不限時間的前提下，每個人至少下兩局，達到14:0的成績<ref name="release"/>，對弈對手包括[[金志錫]]、[[申真谞]]、[[朴永训]]以及[[崔哲瀚]]<ref name="announcement"/>。

== 相關連結 ==
* [[AlphaGo Zero]]，所參考論文的電腦圍棋軟體。
* [[Darkforest]]，上一代[[Facebook]]所開發的電腦圍棋軟體。
* [[Leela Zero]]，另外一套也是依照AlphaGo Zero所實做的[[開源]]電腦圍棋軟體，在ELF OpenGo測試時被當作比較基準。
* [[计算机围棋]]
* [[圍棋軟體]]

== 參考資料 ==
{{Reflist|2}}

== 註解 ==
{{NoteFoot}}

== 外部連結 ==
* {{Official|https://facebook.ai/developers/tools/elf}}
* [https://github.com/pytorch/elf pytorch/ELF: ELF: a platform for game research]{{Wayback|url=https://github.com/pytorch/elf |date=20180503010229 }}

[[Category:围棋软件]]
[[Category:机器学习]]
[[Category:游戏人工智能]]
[[Category:人工智能应用]]
[[Category:使用BSD许可证的软件]]
[[Category:用C++編程的自由軟體]]
[[Category:用Python編程的自由軟體]]
[[Category:用C編程的自由軟體]]

ELF OpenGo - 版本历史

imported>InternetArchiveBot：​补救2个来源，并将0个来源标记为失效。) #IABot (v2.0.9.5

imported>InternetArchiveBot：补救2个来源，并将0个来源标记为失效。) #IABot (v2.0.9.5