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
DRL 組み合わせ最適化
Search
newzy
November 24, 2021
Research
8
90
DRL 組み合わせ最適化
newzy
November 24, 2021
Tweet
Share
Other Decks in Research
See All in Research
Mechanistic Interpretability:解釈可能性研究の新たな潮流
koshiro_aoki
1
410
20250605_新交通システム推進議連_熊本都市圏「車1割削減、渋滞半減、公共交通2倍」から考える地方都市交通政策
trafficbrain
0
770
一人称視点映像解析の最先端(MIRU2025 チュートリアル)
takumayagi
6
3.5k
利用シーンを意識した推薦システム〜SpotifyとAmazonの事例から〜
kuri8ive
1
250
Vision and LanguageからのEmbodied AIとAI for Science
yushiku
PRO
1
530
2021年度-基盤研究B-研究計画調書
trycycle
PRO
0
290
Cross-Media Information Spaces and Architectures
signer
PRO
0
240
MetaEarth: A Generative Foundation Model for Global-Scale Remote Sensing Image Generation
satai
4
190
Learning to (Learn at Test Time): RNNs with Expressive Hidden States
kurita
0
170
MIRU2025 チュートリアル講演「ロボット基盤モデルの最前線」
haraduka
15
7.9k
論文読み会 SNLP2025 Learning Dynamics of LLM Finetuning. In: ICLR 2025
s_mizuki_nlp
0
200
多言語カスタマーインタビューの“壁”を越える~PMと生成AIの共創~ 株式会社ジグザグ 松野 亘
watarumatsuno
0
120
Featured
See All Featured
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
33
2.4k
YesSQL, Process and Tooling at Scale
rocio
173
14k
The Cost Of JavaScript in 2023
addyosmani
53
8.9k
Music & Morning Musume
bryan
46
6.8k
Bootstrapping a Software Product
garrettdimon
PRO
307
110k
For a Future-Friendly Web
brad_frost
180
9.9k
Statistics for Hackers
jakevdp
799
220k
Side Projects
sachag
455
43k
Raft: Consensus for Rubyists
vanstee
140
7.1k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
15
1.6k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
507
140k
Transcript
POMO: Policy Optimization with Multiple Optima for Reinforcement Learning Kwon,
Yeong-Dae, et al. NeurIPS, 2020, vol.33
ཁ •Έ߹Θͤ࠷దԽʹ͓͚ΔɼਂڧԽֶश ͰͷFOEUPFOEͷۙࣅղ๏ɽ •طଘͷਂڧԽֶशख๏ͱൺֱͯ͠ɼ ܭࢉ࣌ؒɾਫ਼ͱʹେ͖͘վળͨ͠ •८ճηʔϧεϚϯͳͲͰݕূɽ 2/26
ಋೖ
Έ߹Θͤ࠷దԽ •८ճηʔϧεϚϯૹܭըɼφοϓβοΫ ͳͲʹද͞ΕΔΑ͏ͳ࠷దͳΈ߹ΘͤΛٻΊΔɽ 4/26 精度 計算時間 厳密解法 最適 遅い 近似解法
最適に 近い 早い https://onl.tw/vzkASMX
ڧԽֶशʢ3FJOGPSDFNFOU-FBSOJOH3-ʣ •3-ɿஞ࣍తͳҙࢥܾఆΛղ͘ख๏ɽ ྦྷੵใु͕࠷େʹͳΔΑ͏ͳํࡦΛݟ͚ͭΔ͜ͱ͕తɽ 5/26 ઃఆͱͯ͠ɼঢ়ଶू߹ɼߦಈू߹ɼใुؔΛ ઃఆ͢Δඞཁ͕͋Δɽ https://onl.tw/98fQVvW
ํࡦϕʔεͷ3&*/'03$& 6/26 •ํࡦ 𝜋 𝑠 ɿঢ়ଶ𝑠ʹ͓͚Δߦಈ𝑎Λग़ྗ͢Δؔ •𝜋! ɿύϥϝʔλ 𝜃ͰύϥϝʔλԽ͞Εͨํࡦ •ํࡦͷߋ৽ࣜɿ𝛼ֶशɼ𝐽
𝜋! తؔ 𝜃 ← 𝜃 + 𝛼∇! 𝐽 𝜋! •ํࡦޯͷࣜɿ𝔼ظɼ𝑅" ऩӹɼ𝑏 𝑠 ϕʔεϥΠϯ ∇! 𝐽 𝜋! = 𝔼#! ∇! log 𝜋! ⋅ 𝑅" − 𝑏 𝑠
ઌߦݚڀ
1PJOUFS/FUXPSLTʢʣ Έ߹Θͤ࠷దԽͰར༻͢ΔωοτϫʔΫ •ॏෳͳ͘બ͠ɼग़ྗύλʔϯྻΛੜ͢Δɽ •ೖྗใ͔Βಛநग़Λߦ͏FODPEFSͱɼFODPEFS ͷग़ྗΛར༻ͯ͑͠ͱͳΔܦ࿏Λग़ྗ͢ΔEFDPEFS͔ ΒͳΔɽ •FODPEFSͱEFDPEFSʹ-45.Λ༻ɽ 8/26
"UUFOUJPO .PEFMʢʣ 1PJOUFS/FUXPSLTͷվྑ൛ •1PJOUFS/FUXPSLTಉ༷ɼ&ODPEFSͱ%FDPEFSΛ༻͢Δ Ϟσϧɽ •-45.ഇࢭ͠ɼ.VMUJIFBE"UUFOUJPOΛ࠾༻ɽ 9/26
ख๏
ຊจͷख๏ͷΞΠσΞ 11/26 ࠷ॳͷߦಈɼޙͷΤʔδΣϯτͷߦಈʹେ͖͘ӨڹΛ༩͑Δɽ Έ߹Θͤ࠷దԽʹΑ͘ݟΒΕΔରশੑΛར༻ɽ
10.0 •3&*/'03$&XJUI#BTFMJOFɿయܕతͳํࡦޯϕʔεͷ 3-ΞϧΰϦζϜΛ༻ɽ •ෳͷҟͳΔ։࢝ߦಈΛࢦఆ͠ɼෳͷߦಈܥྻʢيಓʣ ΛಘΔɽ •ʻ45"35ʼτʔΫϯΛ༻͍ͳ͍ɽ 12/26 従来 POMO
10.0 ∇! 𝐽 𝜃 ≈ 1 𝑁 6 $%& '
𝑅 𝜏$ − 𝑏$ 𝑠 ∇! log 𝑝! 𝜏$ ∣ 𝑠 𝑤ℎ𝑒𝑟𝑒 𝑝! 𝝉$ ∣ 𝑠 ≡ @ "%( ) 𝑝! 𝑎" $ ∣ 𝑠, 𝑎&:"+& $ يಓ 𝝉$ = 𝑎& $ , 𝑎( $ , … , 𝑎) $ GPS 𝑖 = 1,2, … , 𝑁 ڞ༗ϕʔεϥΠϯ 𝑏$(𝑠) = 𝑏TIBSFE (𝑠) = 1 𝑁 6 ,%& ' 𝑅 𝝉, GPS 𝑖 = 1,2, … , 𝑁 13/26
܇࿅෦ͷٖࣅίʔυ 14/26
*OTUBODF"VHNFOUBUJPOɿਪख๏ •ը૾ॲཧͷσʔλΦʔάϝϯςʔγϣϯ͔Βணɽ •ࠓճ͏࠲ඪɼYͷ୯Ґਖ਼ํܗʢୈҰݶʣͷ ͷΛར༻ɽ 15/26 今回使う Instance Augmentation
ਪ෦ͷٖࣅίʔυ 16/26
࣮ݧ
࣮ݧ ࣮ݧ༰ •10.0Λ༻͍ͯɼҎԼͷΛղ͍ͨ݁ՌΛଞͷදతख๏ͱ ൺֱɽ ८ճηʔϧεϚϯ ༰ྔ੍͋Γͷૹܭը φοϓβοΫ
18/26
ֶशۂઢɿ८ճηʔϧεϚϯ 19/26 50地点 100地点
८ճηʔϧεϚϯʢ541ʣ 20/26
८ճηʔϧεϚϯʢ541ʣ 21/26
༰ྔ੍͋Γͷૹܭըʢ$731ʣ 22/26
φοϓβοΫʢ,1ʣ 23/26
࣮ݧͷ·ͱΊ •ҟͳΔઃఆͷͭͷΈ߹Θͤ࠷దԽʹରͯ͠ɼ ಉҰͷ܇࿅ख๏ͱ//ΞʔΩςΫνϟΛ༻͍ͯ༗ͳ݁ՌΛ ಘͨɽ •܇࿅ɾਪख๏ͱͯ͠ͷ10.0ɼਪख๏ͱͯ͠ͷ *OTUBODF"VHNFOUBUJPOͲͪΒޮՌతͳख๏Ͱ͋Δ͜ͱ Λ֬ೝͨ͠ɽ 24/26
·ͱΊ ຊจͰΈ߹Θͤ࠷దԽʹ͓͍ͯɼରশੑΛར༻ ͯ͠3-ͷαϯϓϧޮਫ਼ ਪ࣌ؒΛॖ͢Δख๏Λ հͨ͠ɽ 25/26
ࢀߟจݙ ,XPO :FPOH%BF FUBM10.01PMJDZ0QUJNJ[BUJPOXJUI .VMUJQMF0QUJNBGPS3FJOGPSDFNFOU-FBSOJOH "EWBODFTJO /FVSBM*OGPSNBUJPO1SPDFTTJOH4ZTUFNT
,PPM 8PVUFS )FSLF WBO)PPG BOE.BY8FMMJOH"UUFOUJPO -FBSOUP4PMWF3PVUJOH1SPCMFNT *OUFSOBUJPOBM$POGFSFODF PO-FBSOJOH3FQSFTFOUBUJPOT 7JOZBMT 0SJPM .FJSF 'PSUVOBUP BOE/BWEFFQ+BJUMZ1PJOUFS /FUXPSLT "EWBODFTJO/FVSBM*OGPSNBUJPO1SPDFTTJOH 4ZTUFNT 26/26