Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
入門AlphaGo
Search
na-o-ys
April 22, 2016
Technology
5
3.6k
入門AlphaGo
"Mastering the game of Go with deep neural networks and tree search" の概要
na-o-ys
April 22, 2016
Tweet
Share
More Decks by na-o-ys
See All by na-o-ys
IoTと監視
naoys
1
610
RubyとJIT
naoys
0
130
将棋盤を画像認識したかった
naoys
0
1.4k
Rust で乗り換え案内
naoys
0
590
疎行列と Jaccard 類似度の高速計算
naoys
1
470
有理数集合の濃度
naoys
2
95
YARVの最適化について調べた
naoys
0
100
転職会議サービスのAWS移行記録
naoys
0
30
Anonymous Recursion in C++
naoys
0
390
Other Decks in Technology
See All in Technology
地理空間データ可視化・解析・活用ソリューション Pacific Spatial Solutions (PSS)
pacificspatialsolutions
0
290
KubeCon EU 2024 Recap “Kubernetes Policy Time Machine: Where to Next?”
ryysud
0
220
Google Cloud Next '24でブログを10本書いた方法と勉強会を沸かせた方法
yasumuusan
0
300
require(ESM)とECMAScript仕様
uhyo
3
770
エンジニアのキャリアをちょっと楽しくする3本の軸/Three Pillars to Make an Engineer's Career More Enjoyable
kwappa
0
2.7k
IaCジェネレーターとBedrockで詳細設計書を生成してみた
tsukasa_ishimaru
1
280
アクセス制御にまつわる改善 / Improving access control
itkq
0
550
自己改善からチームを動かす! 「セルフエンジニアリングマネージャー」のすゝめ
shoota
6
780
GraphQL 成熟度モデルの紹介と、プロダクトに当てはめた事例 / GraphQL maturity model
mh4gf
7
1.3k
ServiceNow Knowledge 24の歩き方 EYストラテジー・アンド・コンサルティング
manarobot
0
200
Databricks における 『MLOps』
databricksjapan
2
170
ChatworkのSRE部って実は 半分くらいPlatform Engineering部かもしれない
saramune
0
160
Featured
See All Featured
Unsuck your backbone
ammeep
663
57k
Web development in the modern age
philhawksworth
202
10k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
274
13k
GraphQLの誤解/rethinking-graphql
sonatard
50
9.2k
GitHub's CSS Performance
jonrohan
1025
450k
Building Better People: How to give real-time feedback that sticks.
wjessup
355
18k
4 Signs Your Business is Dying
shpigford
175
21k
Raft: Consensus for Rubyists
vanstee
132
6.3k
A Tale of Four Properties
chriscoyier
151
22k
Happy Clients
brianwarren
92
6.4k
The Art of Programming - Codeland 2020
erikaheidi
42
12k
Making the Leap to Tech Lead
cromwellryan
124
8.5k
Transcript
ೖAlphaGo 0x64ޠ ୈ07 “AI” @na_o_ys
͝ҙ • จʹॻ͔Ε͍ͯͳ͍ಠࣗௐࠪਪଌؚ͕· Ε·͢ • Ұߟͩͱࢥͬͯݟ͍ͯͩ͘͞
AlphaGoͱ • ॳΊͯϓϩع࢜ΛഁͬͨғޟAI
ୈҰ෦: AlphaGoʹࢸΔ·Ͱ
શใήʔϜ • ΦηϩɺνΣεɺকعɺғޟɺetc • ϥϯμϜੑ͕ແ͘ɺ࠷ળख͕ଘࡏ͢Δ • (ݪཧతʹ) ઌखඞউɾޙखඞউɾҾ͖͚
ήʔϜ • શ୳ࡧͰ࠷ળख͕ٻ·Δ • ܭࢉྔతʹෆՄೳ … ݱہ໘ 1खޙ 2खޙ
ධՁؔ • ൫໘Λ༩͑ΔͱείΞ (༧উͳͲ) Λฦؔ͢ • কعνΣεͳΒɺۨͷଛಘޮ͖ͷΛݩʹܭࢉ • ήʔϜͷ୳ࡧൣғ(ਂ͞)ΛݶఆͰ͖Δ ݱہ໘
1खޙ 2खޙ ධՁˠ 0.1 0.8 0.3 0.4
ධՁؔͷ༗ޮੑ • ύϥϝʔλͷબఆɾઃఆ͕ΩϞ • ख࡞ۀ: νΣεͰਓؒΛ͑ͨ • ػցֶश: কعͰਓؒΛ͑ͨ •
ғޟෳࡶੑͷͨΊʹ·ͱͳධՁؔΛ࡞Εͳ͔ͬ ͨ
ݪ࢝ϞϯςΧϧϩ๏ • ϥϯμϜʹऴہ·Ͱଧͭ (ϩʔϧΞτ) Λ܁Γฦͯ͠ɺউΛܭ ࢉ͢Δํ๏ ϩʔϧΞτΛ܁Γฦͯ͠ উΛܭࢉ উ 7/10
উ 3/10
ϞϯςΧϧϩ୳ࡧ (MCTS) • ݪ࢝ϞϯςΧϧϩ๏ΛධՁؔతʹ͏ • n खઌͰϩʔϧΞτ • ༿ͷউΛܭࢉ ※͞Βʹ༿ͷউʹԠͯ͡ಈతʹࢬמΓɾల։͠ɺ୳ࡧਫ਼Λ্͛Δ
ϙϦγʔؔ • f (ہ໘, ࣍ͷҰख) • ࣍ͷҰखͷࣗવ͞Λ͋ΒΘ֬͢ີؔ • عේσʔλ͔Βͷֶश͕༰қ •
ϩʔϧΞτ࣌ʹ͑Δ • ϥϯμϜʹଧͭͷͰͳ͘ɺ·ͱͳखΛଧͨͤΔ • ͨͩ͠ߴʹಈ࡞͢Δඞཁ͕͋Δ
MCTSͷڧ͞ • ϙϦγʔؔͷͳͲͰΞϚνϡΞߴஈʹඖఢ͢Δڧ͞· Ͱਐา • ϓϩʹٴͳ͍ • େہ؍ʹ༏ΕΔ • ʮڱ͘ਂ͍ಡΈʯ͕ऑ͍
• खΛ͘ಡΉͨΊ
AlphaGo͕ͬͨ͜ͱ • جຊMCTS • ༷ʑͳ • CNN(ΈࠐΈχϡʔϥϧωοτϫʔΫ) • ڧԽֶश •
ධՁؔ • ฒྻࢄΞϧΰϦζϜ • MCTS ʹͦΕΒΛΈࠐΜͩ
ୈೋ෦: AlphaGo
2ͭͷϙϦγʔؔͱ 1ͭͷධՁؔ ϩʔϧΞτϙϦγʔ ϩʔϧΞτʹ͏ ߴɾਫ਼ 4-ϙϦγʔ ୳ࡧॱংΛܾΊΔ ɾߴਫ਼ ධՁؔ ༿ͷධՁ(উ)Λܭࢉ
ϩʔϧΞτʹΑΔউͱ͠߹ΘͤΔ
ϩʔϧΞτϙϦγʔ • ϩʔϧΞτ(ϥϯμϜϓϨΠ)ʹ͏ϙϦγʔؔ • ߴੑɹʼɹਫ਼ • ਓؒͷعේ800ສہ໘͔Βֶश • ઢܗιϑτϚοΫεؔ •
2ϚΠΫϩඵ (ߴ) • عේͱͷࢦ͠खҰக: 24.2%
SLϙϦγʔ • ͷ୳ࡧॱংΛܾΊΔϙϦγʔؔ • ਫ਼ɹʼɹߴੑ • ਓؒͷعේ3000ສہ໘͔Βֶश • 13CNN(ΈࠐΈχϡʔϥϧωοτϫʔΫ) •
ը૾ೝࣝͰΑ͘ΘΕΔ • : 3ϛϦඵ • عේͱͷࢦ͠खҰக: 57%
ධՁؔ • 14CNN • SLϙϦγʔΛڧԽֶशͨ͠ͷ (RLϙϦγʔ) Λݩʹɺճؼͯ͠࡞Δ 4-ϙϦγʔ 3-ϙϦγʔ ධՁؔ
1. ڧԽֶश 2. ϥϯμϜعේੜ (3000ສہ໘) 3. ճؼ
ධՁؔͷଊ͑ํ • ϩʔϧΞτʹΑΔউܭࢉΛิ͏ͷ • ୯ମͰͦ͜·Ͱڧ͘ͳ͍ • ධՁؔͷಛ (ߟ) • ʮڱ͘ਂ͍ಡΈʯʹڧ͍
• ʮRLϙϦγʔ(ڧԽֶश݁Ռ)Λऴہ·ͰଧͨͤͨࡍͷউʯͱՁ • େہ؍͕ແ͍ • Ұຊಓ͔͠ಡ·ͳ͍ .$54ͷಛੑ େہ؍ʹ༏Εͯʮਂ͍ಡΈʯ͕ऑ͍ ͱ ͏·͘ิ͍͍͋ͬͯΔ
ڧ͞ (2015/10࣌)
ڧ͞ (2016/3 ࣌) R3500+ ͷΠɾηυϧʹউ
ࢀߟ • Mastering the game of Go with deep neural
networks and tree search (http://www.nature.com/nature/journal/v529/n7587/full/ nature16961.html) • Google AlphaGoͷΈΛཧղ͢Δ | IT Leaders (http://it.impressbm.co.jp/articles/-/13474)
ऴΘΓ