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
TensorFlow & DeepMind Lab & UNREAL
Search
Kosuke Miyoshi
April 20, 2017
Technology
1
2.6k
TensorFlow & DeepMind Lab & UNREAL
TensorFlowで実装したUNREALアルゴリズムでDeepMind Labの3D迷路を解く
Kosuke Miyoshi
April 20, 2017
Tweet
Share
More Decks by Kosuke Miyoshi
See All by Kosuke Miyoshi
Representation Learning with Contrastive Predictive Coding
miyosuda
1
210
Sutton "Reinforcement Learning" 2nd Edition Ch13: Policy Gradient Methods
miyosuda
0
200
Sutton "Reinforcement Learning" 2nd Edition Ch7: n-step Bootstrapping
miyosuda
0
90
Sutton "Reinforcement Learning" 2nd Edition Ch6: TD-learning
miyosuda
0
100
SCAN
miyosuda
0
840
Variational Auto Encoderでの Disentangled表現
miyosuda
0
630
Other Decks in Technology
See All in Technology
Security Hub と出会ってから 1年半が過ぎました
rch850
0
150
会社紹介資料 / Sansan Company Profile
sansan33
PRO
12
400k
習慣とAIと環境 — 技術探求を続ける3つの鍵
azukiazusa1
2
580
善意の活動は、なぜ続かなくなるのか ーふりかえりが"構造を変える判断"になった半年間ー
matsukurou
0
570
Java 25に至る道
skrb
3
230
歴史から学ぶ、Goのメモリ管理基礎
logica0419
14
2.9k
旬のブリと旬の技術で楽しむ AI エージェント設計開発レシピ
chack411
1
290
Oracle Database@AWS:サービス概要のご紹介
oracle4engineer
PRO
2
920
プロンプトエンジニアリングを超えて:自由と統制のあいだでつくる Platform × Context Engineering
yuriemori
0
460
「アウトプット脳からユーザー価値脳へ」がそんなに簡単にできたら苦労しない #RSGT2026
aki_iinuma
11
5.6k
OCI技術資料 : OS管理ハブ 概要
ocise
2
4.1k
サラリーマンソフトウェアエンジニアのキャリア
yuheinakasaka
41
19k
Featured
See All Featured
The browser strikes back
jonoalderson
0
310
Practical Orchestrator
shlominoach
190
11k
Designing for Timeless Needs
cassininazir
0
120
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
133
19k
Lightning talk: Run Django tests with GitHub Actions
sabderemane
0
99
Bash Introduction
62gerente
615
210k
GraphQLの誤解/rethinking-graphql
sonatard
74
11k
How To Speak Unicorn (iThemes Webinar)
marktimemedia
1
370
Claude Code のすすめ
schroneko
67
210k
The World Runs on Bad Software
bkeepers
PRO
72
12k
Imperfection Machines: The Place of Print at Facebook
scottboms
269
13k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
16
1.8k
Transcript
5FOTPS'MPX %FFQNJOE-BC OBSSBUJWFOJHIUTגࣜձࣾ ࡾ߁༞ 5FOTPS'MPX6TFS(SPVQ
%FFQ.JOE-BC
6/3&"- ڧԽֶशͷ"$ΞϧΰϦζϜΛϕʔεʹ&YQFSJFODF 3FQMBZΛͬͨิॿλεΫΛΈ߹Θͤͯ%໎࿏Ͱ YഒͷֶशͷߴԽΛ࣮ݱ REINFORCEMENT LEARNING WITH UNSUPERVISED AUXILIARY TASKS
Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki et. al (DeepMind, 2016)
ಈͷເ w ಈເͷதͰܦݧͨ͠ग़དྷࣄΛ࠶ݱ ϦϓϨΠ ͠ ͳ͕Βւഅ৽ൽ࣭ͷهԱͷݻఆΛߦ͍ͬͯΔ w ߠఆత൱ఆతͳใुʹؔΘΔग़དྷࣄͷເΛಘʹස ൟʹݟֶͯशΛߦ͍ͬͯΔ w
FYʮਫҿΈͰϥΠΦϯΛݟ͔͚ͯةݥͳʹ͋ͬ ͨʯ w 6/3&"-Ͱ͜ΕΛώϯτʹ͍ͯ͠Δ
ڧԽֶश ڥ ΤʔδΣϯτ "DUJPO ⬆ ➡ ⬇ ঢ়ଶ T ใु
S
6/3&"-ͷྲྀΕ %2/ "$ 6/3&"-
"$ "TZODISPOPVT"EWBODFE"DUPS$SJUJD w ෳͷڥΛඇಉظʹฒྻʹಈֶ͔ͯ͠शΛߴԽ ҆ఆԽͤͨ͞
К 1PMJDZ 7 ֤"DUJPOΛऔΔ֬ ݱࡏͷঢ়ଶՁ ⬆ ➡ ⬇ TPGUNBY MJOFBS
$POW $POW '$ -45. "$ͷωοτϫʔΫߏ
֤-PDBM/FUXPSLͰɺֶश݁Ռͷޯ EВ ͷΈΛٻΊɺ ΣΠτʹөͤͣ(MPCBMͷΣΠτ В ʹݸผʹөɻ (MPCBMͷΣΠτΛ·֤ͨ-PDBMͷΣΠτʹίϐʔɻ EВ EВ EВ
EВ В ʜ
1PMJDZ К 7ͷޯ R= = = w 73ʹ͚ۙͮΔ༷ʹߋ৽ w 37͕ਖ਼ͳΒɺऔͬͨBDUJPO͕ग़Δ֬Λ૿༷͢ʹߋ৽
37͕ෛͳΒɺऔͬͨBDUJPO͕ग़Δ֬ΛݮΒ༷͢ʹߋ৽ V network: Policy network: ˞্هͷදهͰ7(SBEJFOU%FTDFOU 1PMJDZ(SBEJFOU"TDFOUθv = θv - α * dθv, θ = θ + α * dθ 1PMJDZ 7
6/3&"- w "$ʹɺ&YQFSJFODF3FQMBZΛޮՌతʹͬͨิ ॿλεΫΛಋೖ͠ɺ͞ΒʹֶशΛߴԽͤ͞Δ w 1JYFM$POUSPM w 3FXBSE1SFEJDUJPO w 7BMVF'VODUJPO3FQMBZ
6/TVQFSWJTFE3&JOGPSDFNFOU"VYJMJBSZ-FBSOJOH
&YQFSJFODF3FQMBZ w <ঢ়ଶ "DUJPO ใु ࣍ঢ়ଶ>ͷϖΞΛେྔʹอଘ͠ ͯɺ͔ͦ͜ΒαϯϓϦϯάͯ͠ωοτϫʔΫΛֶश w %2/ɺ͜Ε͕ͳ͍ͱֶश͕҆ఆ͠ͳ͔ͬͨ w
"$Ͱ͍ͬͯͳ͍
None
1JYFM$POUSPM w ը໘ͷϐΫηϧͷมԽྔΛΑΓେ͖͘͢Δ༷ʹ͞ ͍ͤͨ w ը໘ͷϐΫηϧͷมԽΛٖࣅใुͱ͢Δิॿλε Ϋ
1JYFM$POUSPM w ը໘ΛYͷϐΫηϧάϦουʹ͚ɺάϦουຖʹ2ֶशΛߦ͏ w %VFMJOH/FUXPSLΛͬͨ2ֶश ˞1JYFM$POUSPMͰಘΒΕͨ2͕BDUJPOͷબʹΘΕΔ༁Ͱͳ͍ YͷάϦου BDUJPO ֤άϦουͷϐΫηϧมԽྔฏۉΛใुͱͨ࣌͠ͷׂҾՃࢉใु߹ܭ2
3FXBSE1SFEJDUJPO w &YQFSJFODF3FQMBZ͔Β࿈ଓͨ͠ϑϨʔϜऔΓग़ ͠ɺϑϨʔϜͷใु͕ɺਖ਼͔ෛ͔θϩ͔Λ༧ଌ ͢ΔิॿλεΫ w ༧ଌ͢Δใुɺ ʴ ʔPSͷൺ͕ʹͳΔ༷ʹαϯϓϦϯά ༗ӹͳใुΠϕϯτϨΞͰ͋ͬͯɺසൟʹαϯϓϦϯά͞ΕΔ
3FXBSE1SFEJDUJPO ࣍ͷใु͕ PSPSΛ༧ଌ
7BMVF'VODUJPO3FQMBZ w "$Ͱ͍ͬͯΔɺঢ়ଶՁ 7 ͷਪఆ "DUPS$SJUJDͷ$SJUJDଆ Λɺ&YQFSJFODF3FQMBZ͔ΒαϯϓϦϯάͨ͠ϑϨʔϜͰ࠶ ߦ͏ w 3FXBSE1SFEJDUJPOͱҧͬͯɺαϯϓϦϯάಛʹภΒͤͳ͍
ิॿλεΫɺ"DUJPOબʹӨڹ༩͑ͳ͍͕ɺϕʔ εͷ"$ͱ$POWɺ-45.ͷ8FJHIUΛڞ༗͍ͯ͠Δͷ ͰɺิॿλεΫΛೖΕΔ͜ͱʹΑΓɺͦΕΛղ͘ޮՌతͳ ಛදݱ͕ಘΒΕΔ͜ͱʹΑΓɺؒతʹ"DUJPOબʹӨ ڹΛ༩͑Δ
ଛࣦؔ #BTF"$ 7BMVF'VODUJPO 3FQMBZ 1JYFM$POUSPM YάϦου 3FXBSE 1SFEJDUJPO
None
"$ͱͷൺֱ %FFQ.JOE-BCڥʹͯฏۉͰYഒͷߴԽ
ΓΜ͝ΛऔΔͱ ϫʔϓʹ౸ୡ͢Δͱ ΛಘͯϥϯμϜͳ ॴʹϫʔϓ
࠶ݱݕূಈը IUUQTZPVUVCFY),R#F)* ˞4QFBLFS%FDLͰද͍ࣔͯ͠Δ߹ɺ63-ϦϯΫ͕ΫϦοΫͰ͖ͳ͍ͷͰɺQEGΛμϯϩʔυͯ͠ΫϦοΫ͍ͯͩ͘͠͞
1JYFM$POUSPM ֤άϦουͷલϑϨʔϜͱͷ ϐΫηϧมԽྔ ֤άϦουͷ2 औͬͨ"DUJPOʹର͢Δ2
1PMJDZ К ֤ΞΫγϣϯΛऔΔ֬ લਐ ޙୀ ࠨӈճస ࠨӈεϥΠυ ֶश͕ਐΉͱ΄΅ͷ֬Ͱ֤"DUJPOΛબͿΑ͏ʹͳͬͯ͘Δ
7BMVF'VODUJPO ݱࡏͷঢ়ଶՁ ϫʔϓ ʹۙͮ͘ʹͭΕ্͕͍ͯͬͯ͘
3FXBSE1SFEJDUJPO ϓϥεใु͕དྷΔͱ༧ଌ͍ͯ͠Δ
4PVSDF w IUUQTHJUIVCDPNNJZPTVEBVOSFBM