[DL輪読会]Bridging the Gap Between Value and Policy Based Reinforcement Learning

>100 Views

March 10, 17

#deep learning #Learning #Policy #Algorithm #Implementation #Difficult Interpretation

スライド概要

2017/3/10
Deep Learning JP:
http://deeplearning.jp/seminar-2/

Deep Learning JP

@DeepLearning2023

スライド一覧

DL輪読会資料

またはPlayer版

埋め込む »CMSなどでJSが使えない場合

（ダウンロード不可）

関連スライド

【DL輪読会】KAN: Kolmogorov–Arnold Networks

Deep Learning JP 73.4K

【DL輪読会】Evolutionary Optimization of Model Merging Recipes モデルマージの進化的最適化

Deep Learning JP 53K

【拡散モデル勉強会】拡散モデルの数理

Deep Learning JP 37K

【拡散モデル勉強会】Introduction to Diffusion Models

Deep Learning JP 28.3K

【拡散モデル勉強会】拡散モデルのサンプラーまとめ

Deep Learning JP 24.2K

【DL輪読会】Generative Agents: Interactive Simulacra of Human Behavior

Deep Learning JP 22.9K

各ページのテキスト

#SJEHJOHUIF(BQ#FUXFFO 7BMVFBOE1PMJDZ#BTFE 3FJOGPSDFNFOU-FBSOJOH ॳ୩ྯ࣊

ϝλ৘ใ w (PPHMF#SBJO w w 0pS/BDIVN

3-ΞϧΰϦζϜͷ෼ྨ w 3-ΞϧΰϦζϜ͸େ͖࣍͘ͷ̐ͭʹ෼͚ΒΕΔ w 0OQPMJDZPS0⒎QPMJDZ w 7BMVFCBTFEPS1PMJDZCBTFE

3-BMHPSJUIN`TNBUSJY 7BMVFCBTFE 1PMJDZCBTFE 0OQPMJDZ 0⒎QPMJDZ 4"34" %2/ 3FUSBDF "$ 5310 %%1( "$&3 0⒎1"$

8IBUJT0OQPMJDZ w ‫ࡏݱ‬ͷํࡦͰಘΒΕͨ‫੻ي‬ͷΈΛ࢖ͬͯɹɹɹɹɹ ΤʔδΣϯτΛߋ৽͢Δ w ‫ࡏݱ‬ͷํࡦКʹґଘͨ͠2΍7Λ‫ٻ‬ΊΔࣄ͕ଟ͍

3-BMHPSJUIN`TNBUSJY 7BMVFCBTFE 1PMJDZCBTFE 0OQPMJDZ 0⒎QPMJDZ 4"34" %2/ 3FUSBDF "$ 5310 %%1( "$&3 0⒎1"$

4"34" Q π ํࡦΛКʹ‫ݻ‬ఆͨ͠ͱ͖ͷ2ؔ਺ #FMMNBO&RVBUJPO π π Q (s,a) = r(s,a) + γ Eπ [Q (s ',a')] NJOJNJ[F-GPSВ π θ π θ L = (r(s,a) + γ Q (s ',a') − Q (s,a)) 2

8IBUJT0⒎QPMJDZ w ‫ࡏݱ‬ͷํࡦͱ͸ҧ͏ํࡦ͔ΒಘΒΕͨ‫੻ي‬΋࢖ͬͯ ΤʔδΣϯτΛߋ৽͢Δ w ࠷దํࡦʹର͢Δ2΍7Λ௚઀‫ٻ‬ΊΔ͜ͱ͕ଟ͍

3-BMHPSJUIN`TNBUSJY 7BMVFCBTFE 1PMJDZCBTFE 0OQPMJDZ 0⒎QPMJDZ 4"34" %2/ 3FUSBDF "$ 5310 %%1( "$&3 0⒎1"$

10.

2MFBSOJOH %2/ o Q ࠷దํࡦʹର͢Δ2ؔ਺ #FMMNBO&RVBUJPO Q (s,a) = r(s,a) + γ max Q (s ',a') o o a' NJOJNJ[F-GPSВ L = (r(s,a) + γ max Q (s ',a') − Q (s,a)) a' o θ o θ 2

11.

0OQPMJDZWT0⒎QPMJDZ w w 0OQPMJDZNFUIPE ๏ ߋ৽ࣜʹϚϧνεςοϓΛ͙͢ద༻Ͱ͖Δ º ֶशͷͨΊͷ‫੻ي‬Λ౎౓αϯϓϧ͠ͳ͍ͱ͍͚ͳ͍ 0⒎QPMJDZNFUIPE ๏ ࠓ·Ͱಘͨ‫੻ي‬Λશֶͯशʹར༻Ͱ͖Δ º NBYPQFSBUPSͷ͍ͤͰ̍εςοϓ͔͠ߋ৽ʹ࢖͑ͳ͍

12.

13.

/TUFQ4"34" #FMMNBO&RVBUJPO π π Q (s,a) = r(s,a) + γ Eπ [Q (s ',a')] OTUFQ෼ͷ‫੻ي‬Λ༻͍Δ n−1 L = (∑ γ r(si ,ai ) + γ Q (sn ,an ) − Q (s0 ,a0 )) i n π θ π θ 2 i=0 TUFQΛେ͖͘͢Δ͜ͱͰUBSHFUਪఆͷCJBTΛ‫ݮ‬Β͢

14.

15.

2MFBSOJOH %2/ L = (r(s,a) + γ max Q (s ',a') − Q (s,a)) a' o θ o θ NBYPQFSBUPSͰબ୒͞ΕͨB`ͱ ࣮ࡍͷ‫੻ي‬ͷB`͕ҟͳΔͷͰ NVMUJTUFQʹͰ͖ͳ͍ 2

16.

7BMVFCBTFEͱ1PMJDZCBTFE w w 7BMVFCBTFE ‎ ࠷దͳՁ஋ؔ਺Λ‫ٻ‬ΊΔ ˒ ֶशํ๏Ձ஋ؔ਺ͷpUUJOH 1PMJDZCBTFE ‎ ࠷దͳํࡦΛ௚઀‫ٻ‬ΊΔ ˒ ֶशํ๏ํࡦޯ഑๏

17.

$POUSJCVUJPO w FOUSPQZSFHVMBSJ[FEͳQPMJDZʹؔͯ͠ɹɹɹɹɹ ҰൠԽͨ͠ϕϧϚϯ࠷దํఔࣜΛఏҊ w ͦΕΛ‫ʹݩ‬P⒎QPMJDZͰNVMUJTUFQͳΞϧΰϦζϜ 1$- 1BUI$POTJTUFODZ-FBSOJOH ΛఏҊ w 7BMVFCBTFEͱ1PMJDZCBTFEͳख๏Λ౷Ұతͳɹ ‫Ͱํݟ‬આ໌

18.

&OUSPQZSFHVMBSJ[FEͱ͸ w ํࡦͷ෼෍͕POFIPUʹͳΒͳ͍Α͏ʹํࡦͷ FOUSPQZΛ࠷େԽͤ͞ͳ͕Βֶशͤ͞Δ w "$ͳͲͰ͸ଛࣦؔ਺ͷ̍෦෼Ͱ࢖ΘΕ͍ͯΔ

19.

(FOFSBMCFMMNBOFRVBUJPO ∗ Q FOUSPQZSFHVMBSJ[FEͳ࠷ద2ؔ਺ #FMMNBO&RVBUJPO Q (s,a) = r(s,a) + γτ log ∑ a' exp(Q (s ',a') / τ ) ∗ ∗ w ϕϧϚϯ࠷దํఔࣜ಺ͷNBYPQFSBUPSΛɹɹɹɹ MPHTVNFYQʹ͢Δ͜ͱͰҰൠԽ w НˠͰNBYPQFSBUPSͱҰக

20.

Нˠͷ࣌ τ log ∑ a' exp(Q (s ',a') / τ ) ∗ = τ log(exp(Q (s ,a ) / τ )∑ a' exp((Q (s ',a') − Q (s ,a )) / τ )) ∗ M ∗ M ∗ M M = max Q (s ',a') + τ log( ∑ a' exp((Q (s ',a') − Q (s ,a )) / τ )) ∗ ∗ ∗ M M a' ҎԼ w "QQFOEJYʹ͸FOUSPQZSFHVMBSJ[FEͳํࡦʹɹɹ ͜ͷํఔ͕ࣜ੒Γཱͭ͜ͱ͕ࣔ͞Ε͍ͯΔ

21.

Ϛϧνεςοϓͷલʹ w ͋Δ̍ͭͷεςοϓͰͷКʹؔ͢Δ໨తؔ਺ͷ ࠷େԽͰ‫ࡏݱ‬ͷ࠷దঢ়ଶՁ஋ؔ਺ͷਪఆΛߦ͏ w ཁ͸#FMMNBOํఔࣜͷಋग़Ͱ΍͍ͬͯΔ͜ͱΛɹɹ ໨తؔ਺ͷ࠷େԽͱ͍͏จ຺Ͱߟ͑Δʁ ҎԼFOUSPQZSFHVMBSJ[FEͳ7ͱКʹ͍ͭͯ V (s) = −τ log π (a | s) + r(s,a) + γ V (s ') ∗ ∗ ͕੒ཱ͢Δ͜ͱʹ͍ͭͯઆ໌͠·͢ ৄ͍͠ূ໌͸࿦จͷ"QQFOEJYʹ ∗

22.

࠷దঢ়ଶՁ஋ؔ਺ͷ࣌ {a1 ,...,an } ࣍ͷঢ়ଶ ‫ࡏݱ‬ͷঢ়ଶ s0 ,v0 ໨తؔ਺ OMR (π ) = {v1 ,...,vn } {s1 ,..., sn } n ∑ π (a )(r + γ v ) i o i i i=1 ໨తؔ਺Λ࠷େԽͤ͞Δ࣌К͸POFIPUʹͳΔ v = OMR (π ) = max(ri + γ v ) o 0 o i o i

23.

&OUSPQZSFHVMBSJ[FEͳ࣌ᶃ n ໨తؔ਺ OENT (π ) = ∑ π (ai )(ri + γ v − τ log π (ai )) ∗ i i=1 π (ai ) OENT (π ) = −τ ∑ π (ai )log +τS ∗ exp((ri + γ vi ) / τ ) / S i=1 n ໨తؔ਺ͷ࠷େԽ ˠ,-μΠόʔδΣϯεͷ࠷খԽ π (ai ) = ∗ exp((ri + γ v ) / τ ) ∗ i n ∑ exp((r i' i '=1 +γ v ) /τ ) ∗ i' ࣜม‫ܗ‬

24.

&OUSPQZSFHVMBSJ[FEͳ࣌ᶄ n v = OENT (π ) = τ log ∑ exp((ri + γ v ) / τ ) ∗ 0 ∗ ∗ i i=1 π (ai ) = exp((ri + γ v ) / τ ) ∗ i ∗ n ∑ exp((r i' +γ v ) /τ ) ∗ i' i '=1 v = −τ log π (ai ) + r(si ,ai ) + γ v ∗ 0 ∗ ∗ i

25.

$POTJTUFODZ w ҰൠʹҎԼͷ౳͕ࣜશͯͷ T B S Ͱ੒ཱ͢Δɹɹɹɹɹɹ ূ໌͸"QQFOEJY V (s) = −τ log π (a | s) + r(s,a) + γ V (s ') ∗ ∗ ∗ ‫ؼ‬ೲతʹద༻ −V (s1 ) + γ V (st ) + R(s1:t ) − τ G(s1:t , π ) = 0 ∗ t−1 n−m−1 ∗ ∗ n−m−1 R(sm:n ) = ∑ γ r(sm+i ,am+i ) G(sm:n , π ) = ∑ γ log π (am+i | sm+i ) i i=0 i i=0

26.

1$QBSBNFUFSJ[F Cθ ,φ (s1:t ) = −Vφ (s1 ) + γ Vφ (st ) + R(s1:t ) − τ G(s1:t , π θ ) t−1 w $ͷೋ৐Λଛࣦؔ਺ͱͯ͠༻͍Δ͜ͱͰВͱП΁ͷ ࠷దԽ໰୊ͱͯ͠‫ؼ‬ண͢Δ͜ͱ͕Ͱ͖Δ w QPMJDZCBTFEͱWBMVFCBTFEΛ౷ҰతʹఆࣜԽ Δθ ∝ Cθ ,φ (s1:t )∇θ G(s1:t , π θ ) Δφ ∝ Cθ ,φ (s1:t )(∇φVφ (s1 ) − ∇φγ Vφ (st )) t−1

27.

28.

"$ͱͷൺֱ $POTJTUFODZ Cθ ,φ (s1:t ) = −Vφ (s1 ) + γ Vφ (st ) + R(s1:t ) − τ G(s1:t , π θ ) t−1 "$ߋ৽ࣜ Aθ ,φ (s1:d+1 ) = −Vφ (s1 ) + γ Vφ (sd+1 ) + R(s 1:d+1 ) d T −1 Δθ ∝ Es0:T [∑ Aθ ,φ (si:i+d )∇θ log π θ (ai | si )] i=0 T −1 Δφ ∝ Es0:T [∑ Aθ ,φ (si:i+d )∇φVφ (si )] i=0

29.

%2/ͱͷൺֱ w NBYPQFSBUPSʹΑͬͯ̍εςοϓ͔͠ߋ৽ࣜʹɹ ૊ΈࠐΊͳ͔ͬͨ΋ͷΛϚϧνεςοϓʹ֦ு w ࣮‫ݧ‬తʹ্ख͘ߦ͍ͬͯͨOTUFQ2ʹMPHQSPCΛ ଍͢͜ͱͰཧ࿦తʹ‫ͮ͘ج‬ΞϧΰϦζϜͱ֦ͯ͠ு

30.

࣮‫ݧ‬ w ΞϧΰϦζϜ‫ܥ‬λεΫ w "$ͱ1SJPSJUJ[FE%%2/Ͱൺֱ w શͯͷλεΫͰ%2/ "$ʹಉ౳΋͘͠͸উར

31.

32.

࣮‫ݧ‬ΤΩεύʔτ w ΤΩεύʔτͷ‫੻ي‬Λ3#ʹೖΕֶͯश w w JNQPSUBODFTBNQMJOHΛ࢖͏ख๏ͱҧ͍෼෍͕ Θ͔Βͳͯ͘΋࢖༻Մೳ ΋ͷ͘͢͝ྑ͘ͳͬͨ

33.

34.

·ͱΊ w 0⒎QPMJDZͰϚϧνεςοϓͳֶश͕Մೳ w Ձ஋ؔ਺ۙࣅͱํࡦޯ഑Λ౷Ұͨ͠ଛࣦؔ਺ͷɹɹ ࠷খԽͰఆࣜԽͨ͠ w ‫ط‬ଘख๏ΛPWFSQFSGPSNͨ͠

35.

‫ݸ‬ਓతʹࢥͬͨ͜ͱ w ࣮‫͕ݧ‬τΠλεΫతͳ΋ͷ͔͠΍͍ͬͯͳ͍ͷ͸ͳ ͥʁ ‎ $BSU1PMF΍ͬͯΈͨ w w IUUQTHJUIVCDPNSBSJMVSFMPQDM@LFSBT $POUJOVPVTDPOUSPMλεΫʹ΋ߋ৽ࣜ͸ͦͷ·· ࢖͑ͦ͏Ͱ ূ໌͸೉ͦ͠͏ ࢼͯ͠ΈΔՁ஋͕͋Γ ͦ͏

36.

$BSU1PMF