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
PRML第11章
Search
gucchi
March 30, 2019
Science
360
1
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
PRML第11章
gucchi
March 30, 2019
More Decks by gucchi
See All by gucchi
PRML(ニューラルネット編)
gucchi
1
340
PRML(分類編)
gucchi
2
510
PRML(回帰編)
gucchi
2
600
PRML第10章
gucchi
1
350
PRMLセミナー(第9章)
gucchi
3
430
PRMLセミナー
gucchi
2
330
PRMLセミナー
gucchi
1
410
PRMLセミナー
gucchi
1
600
PRML第6章
gucchi
1
67
Other Decks in Science
See All in Science
Physical AIを支えるWeights & Biases
olachinkei
1
350
HajimetenoLT vol.17
hashimoto_kei
1
240
Cross-Media Technologies, Information Science and Human-Information Interaction
signer
PRO
3
32k
データベース05: SQL(2/3) 結合質問
trycycle
PRO
0
1.2k
データベース06: SQL (3/3) 副問い合わせ
trycycle
PRO
1
970
見上公一.pdf
genomethica
0
140
データベース09: 実体関連モデル上の一貫性制約
trycycle
PRO
0
1.2k
データベース03: 関係データモデル
trycycle
PRO
1
540
ハミルトン・ヤコビ方程式の解の性質と物理的意味
enakai00
0
610
AI(人工知能)の過去・現在・未来 —AIは人間を超えるのか—
tagtag
PRO
0
120
(メタ)科学コミュニケーターからみたAI for Scienceの同床異夢
rmaruy
0
240
Question Driven Development using Python
willingc
PRO
1
120
Featured
See All Featured
Imperfection Machines: The Place of Print at Facebook
scottboms
270
14k
Why You Should Never Use an ORM
jnunemaker
PRO
61
9.9k
Writing Fast Ruby
sferik
630
63k
Principles of Awesome APIs and How to Build Them.
keavy
128
17k
Ecommerce SEO: The Keys for Success Now & Beyond - #SERPConf2024
aleyda
1
2k
Game over? The fight for quality and originality in the time of robots
wayneb77
1
190
技術選定の審美眼(2025年版) / Understanding the Spiral of Technologies 2025 edition
twada
PRO
118
120k
Taking LLMs out of the black box: A practical guide to human-in-the-loop distillation
inesmontani
PRO
3
2.3k
AI Search: Where Are We & What Can We Do About It?
aleyda
0
7.6k
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
162
16k
Creating an realtime collaboration tool: Agile Flush - .NET Oxford
marcduiker
35
2.5k
AI: The stuff that nobody shows you
jnunemaker
PRO
8
690
Transcript
PRML ୈ 11 ষ αϯϓϦϯά๏ 2019/04/01 ࡔޱ ྒี 1 /
56
Ұൠతͳ֬Ϟσϧʹ͓͍ͯɺ࿈ଓతͳ֬มΛ z ͱͨ͠ͱ͖ͷ͋Δ ؔ f(z) ͷ֬ p(z) ͷԼͰͷظ E[f] =
∫ f(z)p(z) dz (11.1) Λܭࢉ͍ͨ͠γνϡΤʔγϣϯ͕ଟ͘ൃੜ͢Δɻ(ྫ͑ɺPRML ͷ 4 ষͷࣜ (4.145) ͳͲ) ଟ͘ͷ߹ɺ͜ͷੵ (11.1) ղੳతʹܭࢉͰ͖ͳ͍ɻ αϯϓϦϯά๏ͷҰൠతͳΞΠσΞɺ p(z) ͔Βಠཱʹநग़͞Ε ͨαϯϓϧ z(l) (l = 1, · · · , L) Λಘͯɺ(11.1) ΛҎԼͷΑ͏ʹۙࣅ͢Δɻ f = 1 L L ∑ l=1 f(z(l)) (11.2) ͔͜͠͠ͷํ๏ͩͱɺ͠ p(z) ͕খ͍͞ྖҬͰ f(z) ͕େ͖͘ɺp(z) ͕ େ͖͍ྖҬͰ f(z) ͕খ͔ͬͨ͞Βɺαϯϓϧ͕গͳ͍࣌ͦͷαϯ ϓϧͷதʹ p(z) ͕খ͍͞ྖҬ͔Βαϯϓϧ͕͋ͬͨΒɺ(11.2) p(z) ͕খ͍͞ྖҬ͔ΒαϯϓϧʹӨڹΛड͚͗͢Δͱ͍͏͕͋Δɻ 2 / 56
11.1 جຊతͳαϯϓϦϯάΞϧΰϦζϜ ͜͜Ͱɺ༩͑ΒΕ͔ͨΒϥϯμϜʹαϯϓϦϯάΛߦ͏ํ๏Λઆ ໌͢Δɻ αϯϓϧࣗମܭࢉػʹΑΓੜ͞Ε͍ͯΔͨΊɺαϯϓϧٖࣅཚ Ͱ͋Δɻ ͭ·ΓɺܾఆతͳܭࢉͷΈ߹ΘͤʹΑΓɺϥϯμϜͳαϯϓϧΛੜ ͢Δɻ ·ͨԾఆͱͯ͠ɺ۠ؒ (0,
1) ͷҰ༷͔ΒͷαϯϓϧੜΛߦ͏Ξϧ ΰϦζϜ༩͑ΒΕ͍ͯΔͱ͢Δɻ ͜ͷҰ༷͔ΒͷαϯϓϦϯά͔ΒඇҰ༷ͷαϯϓϦϯάΛ ߦ͏ɻ 3 / 56
11.1.1 ඪ४తͳ ·ͣɺz ͕۠ؒ (0, 1) ͰҰ༷ʹ͓ͯ͠Γɺͦͷม z Λ y
= f(z) Ͱ ඇҰ༷ͳ֬ม y ʹม͢Δ͜ͱΛߟ͑Δɻ ͜ͷͱ͖ɺPRML ͷࣜ (1.27) ΑΓɺy ͷ֬ҎԼͷΑ͏ʹͳΔɻ p(y) = p(z) dz dy (11.5) ͜͜Ͱɺp(z) ۠ؒ (0, 1) ͷҰ༷ͳͷͰɺp(z) = 1 ͱͳΔɻ (p(z) = c = const. ͱͯ͠ɺن֨Խ݅Λ՝ͤɺc = 1 ͱͳΔɻ) ͜͜Ͱɺy Λ z ʹม͢ΔҎԼͷؔ h(y) Λߟ͑Δɻ z = h(y) ≡ ∫ y −∞ p(ˆ y) dˆ y (11.6) ͜ͷΑ͏ͳؔ h(y) Ͱ z ʹม͢Εɺp(z) = 1 ͱ h(y)′ = p(y) ≥ 0 Λ༻͍ͯ p(z) dz dy = |h(y)′| = p(y) ͱͳΓɺ֬ม y p(y) ʹै͏͜ͱ͕Θ͔Δɻ 4 / 56
11.1.1 ඪ४తͳ ͜ΕΑΓɺp(y) ʹै͏ y Λ z ͔ΒٻΊΔʹɺ(11.6) Ͱఆٛ͞Εͨؔ h
ͷٯؔΛ༻͍ͯɺy = h−1(z) ͱͯ͠ม͢Εྑ͍ɻ ͳͷͰɺҰ༷͔ΒಘΒΕͨαϯϓϧ zi ͔Βɺ͋ΔඇҰ༷ p(y) ͔Β༩͑ΒΕΔαϯϓϧ yi ΛಘΔͨΊͷखॱҎԼͷΑ͏ʹͳΔɻ 1. (11.6) ΑΓɺp(y) Λੵͯ͠ h(y) ΛಘΔɻ 2. h(y) ͷٯؔ h−1(z) ΛٻΊΔɻ 3. h−1 Λ༻͍ͯɺyi = h−1(zi ) Ͱαϯϓϧ yi ΛಘΔɻ ͜ͷखॱͰͷαϯϓϦϯά๏Λٯؔ๏ͱ͍͏ɻ 5 / 56
11.1.1 ඪ४తͳ ྫͱͯ͠ɺαϯϓϧΛҎԼͷࢦ͔Βಘ͍ͨͱ͢Δɻ p(y) = λ exp (−λy) (11.7) ͜͜Ͱɺλ
ਖ਼ͷύϥϝʔλͰɺ0 ≤ y < ∞ Ͱ͋Δɻ ·ͣ (11.6) ΑΓɺ0 ≤ y < ∞ Ͱ͋Δ͜ͱʹҙͯ͠ɺ h(y) = ∫ y 0 λ exp (−λˆ y) dˆ y = − [ exp (−λˆ y) ]ˆ y=y ˆ y=0 = 1 − exp (−λy) ͱͳΓɺh(y) = z ͱ͢Δͱɺy (−λy ≤ 0 ΑΓɺ1 − exp (−λy) < 1 Ͱ ͋Δ͜ͱʹҙ͢Δͱ) z = 1 − exp (−λy) → − λy = ln (1 − z) →y = −λ−1 ln (1 − z) ͱͳΔɻ 6 / 56
11.1.1 ඪ४తͳ ͜ΕΑΓɺҰ༷͔Β zi ΛಘΔͱɺyi = −λ−1 ln (1 −
zi ) ͱͯ͠ಘͨ yi ࢦ (11.7) ͔Βͷαϯϓϧʹͳ͍ͬͯΔɻ ͜ͷํ๏͕͏·͍͘͘ʹɺੵ (11.6) ͕࣮ߦͰ͖ɺٯؔ h−1 ͕ٻ ·Δඞཁ͕͋Δɻ ্ͷྫͷࢦͷΑ͏ͳ୯७ͳͰ͋ΕՄೳͰ͋Δ͕ɺҰൠతʹ ෆՄೳͰ͋Δɻ ͦͷΑ͏ͳෳࡶͳʹ͍ͭͯผͷΞϓϩʔν͕ඞཁͰɺ࣍ʹड़ Δغ٫αϯϓϦϯάॏαϯϓϦϯά͕༗ޮͰ͋Δɻ 7 / 56
11.1.2 غ٫αϯϓϦϯά ࣍ͷغ٫αϯϓϦϯάͷઆ໌Λߦ͏ɻ ͜͜Ͱɺલͱಉ͡Α͏ʹ p(z) ͔ΒαϯϓϦϯάΛߦ͍͍ͨͱ ͢Δɻ ·ͨԾఆͱͯ͠ɺp(z) ن֨ԽఆΛআ͍ͯΘ͔͍ͬͯΔͱ͢Δɻ ͭ·Γɺن֨ԽఆΛ
Zp ͱͯ͠ɺp(z) Λ p(z) = 1 Zp p(z) (11.13) ͱ͢Δͱɺp(z) Θ͔͍ͬͯΔ͕ɺZp ͷΘ͔͍ͬͯͳ͍ͱ͢Δɻ غ٫αϯϓϦϯάΛߦ͏ʹ͋ͨͬͯɺΑΓαϯϓϦϯά͕؆୯ͳ (11.1.1 ͷٯؔ๏ͰαϯϓϦϯάͰ͖ΔΑ͏ͳ) q(z) Λ༻ҙ͢Δɻ (͜ͷΑ͏ͳΛఏҊͱ͍͏ɻ) 8 / 56
11.1.2 غ٫αϯϓϦϯά ࣍ʹɺਖ਼ͷఆ k Λ༻ҙͯ͠ɺͯ͢ͷ z ʹରͯ͠ kq(z) ≥ p(z)
ͱͳ ΔΑ͏ʹ k ͷΛܾΊΔɻ p(z) Θ͔͍ͬͯΔ͔Βɺk ͷٻΊΒΕΔɻ ҎԼ͕ kq(z) ͱ p(z) ͷྫͰ͋Δɻ 9 / 56
11.1.2 غ٫αϯϓϦϯά ࣮ࡍͷαϯϓϦϯάͷํ๏ɺ·ͣఏҊ q(z) ͔Βཚ z0 Λੜ ͢Δɻ ࣍ʹ۠ؒ [0,
kq(z0 )] ͷҰ༷͔Βཚ u0 Λੜ͢Δɻ ੜͨ͠ u0 ͕ɺu0 > p(z0 ) Ͱ͋Εغ٫͞Ε (ࣺͯΒΕ)ɺͦΕҎ֎Ͱ ͋Ε u0 p(z) ͔ΒͷαϯϓϦϯάͱͯ͠อ࣋͞ΕΔɻ 10 / 56
11.1.2 غ٫αϯϓϦϯά αϯϓϦϯάͰ͖Δ͚ͩޮతʹߦ͍͍ͨͷͰɺغ٫͞ΕΔαϯϓϧ ͷݮΒ͍ͨ͠ɻ ͦ͜Ͱɺαϯϓϧ z ͕डཧ͞ΕΔ֬ p(accept) ΛٻΊͯΈΔɻ αϯϓϧ
z q(z) ͔Βੜ͞Εɺ۠ؒ [0, kq(z)] ͷҰ༷͔Βੜ͞ ΕΔαϯϓϧ p(z)/kq(z) ͷ֬Ͱडཧ͞ΕΔͷͰɺαϯϓϧ͕डཧ ͞ΕΔ֬ p(accept) p(accept) = ∫ {p(z)/kq(z)}q(z) dz = 1 k ∫ p(z) dz = ∫ p(z) dz ∫ kq(z) dz (11.14) ͱͳΔɻ 11 / 56
11.1.2 غ٫αϯϓϦϯά ͭ·Γɺ(11.14) ΑΓɺk ͯ͢ͷ z ʹରͯ͠ kq(z) ≥ p(z)
Λຬͨ͢ ݶΓɺͰ͖Δ͚ͩখ͘͢͞Δඞཁ͕͋Δɻ ·ͨɺԼͷਤͷփ৭ͷ෦ΛͰ͖Δ͚ͩখ͘͢͞ΔΑ͏ͳఏҊΛબ ɺडཧ্͕͕Δɻ 12 / 56
11.1.3 దԠతغ٫αϯϓϦϯά غ٫αϯϓϦϯάͰɺదͳఏҊΛ༻ҙ͢Δඞཁ͕͋Δ͕ɺద ͳఏҊ͕Θ͔Βͳ͍߹ɺదԠతغ٫αϯϓϦϯά͕͑Δɻ ͜ͷαϯϓϦϯάͰɺαϯϓϦϯάΛ͍ͨ͠ p(z) Θ͔͍ͬͯ Δͱ͢Δɻ ͞ΒʹɺҎԼͷ͍ؔΑ͏ʹ p(z)
ରԜ (ର্͕ʹತͳؔ) Ͱ ͋Δͱ͢Δɻ 13 / 56
11.1.3 దԠతغ٫αϯϓϦϯά ·ͣॳظू߹ͱͯ͠ɺԿ͔αϯϓϧ {zi } ͕༩͑Δɻ ͦͯ͠ɺͦͷͰͷ ln p(z) ͷඍΛ
−λi ͱ͢Δɻ d dz ln p(z) z=zi = −λi ͜ͷඍ −λi Λ༻͍ͯɺఏҊͷର ln q(z) ΛҎԼͷΑ͏ʹ ࡞Δɻ ln q(z) = −λi (z − zi ) + ln λi ki (ˆ zi−1,i < z ≤ ˆ zi,i+1 ) ͜͜Ͱɺˆ zi−1,i zi−1 Ͱͷઢͱ zi Ͱͷઢͱͷަͷ z ࠲ඪɻ 14 / 56
11.1.3 దԠతغ٫αϯϓϦϯά ͜ͷΑ͏ʹͯ͠࡞ΒΕͨఏҊ p(z) ͕ରԜͰ͋Εɺ p(z) ≤ q(z) ͱͳΓɺغ٫αϯϓϦϯά͕ద༻Ͱ͖Δɻ ͜ΕΑΓɺq(z)
q(z) = λi ki exp {−λi (z − zi )} (ˆ zi−1,i < z ≤ ˆ zi,i+1 ) ͱͳΔɻ ͜ͷ q(z) ͔ΒͷαϯϓϦϯά؆୯Ͱɺ11.1.1 ͷٯؔ๏Λ༻͍ͯα ϯϓϦϯάͰ͖Δɻ(ԋश 11.9) ͦ͜Ͱɺq(z) ͔Βͷ৽ͨͳαϯϓϦϯά z ͕༩͑ΒΕͨΒɺغ٫αϯϓ Ϧϯάͷͱ͖ͱಉ༷ʹ [0, q(z)] ͷҰ༷͔Βαϯϓϧ u ΛಘΔɻ ͦͯ͠ɺu > p(u) ͳΒغ٫͠ɺͦΕҎ֎ͳΒडཧ͢Δɻ غ٫͞ΕͨΒɺq(z) Λ࡞͢ΔͨΊͷ৽͍͠ʹ͢Δɻ 15 / 56
11.1.3 దԠతغ٫αϯϓϦϯά ͜͜Ͱɺغ٫αϯϓϦϯά͕αϯϓϧม z ͕ߴ࣍ݩϕΫτϧͷ࣌ʹ ͔ͳ͍͜ͱΛઆ໌͢Δɻ ྫͱͯ͠ɺฏۉ͕θϩͰڞࢄߦྻ͕ σ2 p I
Ͱ͋ΔΨε͔Βαϯϓ Ϧϯά͍ͨ͠ͱ͢Δɻ p(z) = N(z|0, σ2 p I) ·ͨɺఏҊ q(z) ฏۉ͕θϩͰڞࢄߦྻ͕ σ2 q I Ͱ͋ΔΨε Ͱ͋Δͱ͢Δɻ q(z) = N(z|0, σ2 q I) غ٫αϯϓϦϯάͰɺͯ͢ͷ z Ͱ kq(z) ≥ p(z) Ͱ͋ΔΑ͏ͳ k ͕ ଘࡏ͠ͳ͍͚ͯ͘ͳ͍ɻ ͦͷΑ͏ͳ k ͕ଘࡏ͢Δʹɺq(z) ͕ p(z) ΑΓฏ͍ͨͰͳ͍ͱ͍ ͚ͳ͍ɻ ͭ·Γɺσ2 q ≥ σ2 p Ͱ͋Δඞཁ͕͋Δɻ 16 / 56
11.1.3 దԠతغ٫αϯϓϦϯά σ2 q ≥ σ2 p ͷ݅Ͱɺͯ͢ͷ z Ͱ
kq(z) ≥ p(z) ͱͳΔͨΊʹҎԼͷਤ ͷΑ͏ʹɺ࠷େΛ༩͑ΔͰ z = 0 Ͱ kq(z = 0) = p(z = 0) ͱͳΕ Α͘ɺ N(x|µ, Σ) = 1 (2π)D/2 1 |Σ|1/2 exp { − 1 2 (x − µ)TΣ−1(x − µ) } ΑΓɺ k = ( σq σp )D ͱऔΕྑ͍ɻ 17 / 56
11.1.3 దԠతغ٫αϯϓϦϯά ͜ΕΑΓɺαϯϓϦϯάͷडཧ p(accept) (11.14) ΑΓ p(accept) = ∫
p(z) dz ∫ kq(z) dz = 1 k = ( σp σq )D ͱͳΔɻ ͭ·ΓɺD = 1000 Ͱɺσq /σp = 1.01 ͷͱ͖ (σq ͕ͨͬͨ 1 ύʔηϯ τ͚ͩ σp ΑΓେ͖͍ͱ͖) p(accept) = ( 1 1.01 )1000 ∼ 1 20000 ͱͳΓɺ΄ͱΜͲडཧ͞Εͳ͍ɻ ͭ·Γɺغ٫αϯϓϦϯάߴ࣍ݩʹద͓ͯ͠Βͣɺ1 ࣍ݩ·ͨ 2 ࣍ݩ͘Β͍ͷͱ͖ʹదͨ͠αϯϓϦϯάͰ͋Δɻ 18 / 56
11.1.4 ॏαϯϓϦϯά ॏαϯϓϦϯάͰɺ p(z) ͔ΒαϯϓϦϯάΛಘΔͷͰͳ͘ɺ (11.1) ͷΑ͏ͳظͷۙࣅΛٻΊΔɻ (11.2) ͷΑ͏ʹۙࣅ͢Δͱɺ{z(l)} ෳࡶͳ
p(z) ͔Βͷαϯϓϧͳ ͷͰɺαϯϓϦϯά͕͍͠ɻ ͦ͜Ͱɺ͜͜ͰαϯϓϦϯά͕ൺֱత؆୯ͳఏҊ q(z) Λར༻ ͢Δɻ ఏҊΛར༻ͯ͠ɺ(11.1) ΛҎԼͷΑ͏ʹۙࣅ͢Δɻ E[f] = ∫ f(z)p(z) dz = ∫ f(z) p(z) q(z) q(z) dz ∼ 1 L L ∑ l=1 p(z(l)) q(z(l)) f(z(l)) (11.19) ͜͜Ͱɺ{z(l)} q(z) ͔ΒͷαϯϓϧͰ͋Δɻ 19 / 56
11.1.4 ॏαϯϓϦϯά ͜͜Ͱɺrl = p(z(l))/q(z(l)) ॏཁॏΈͱ͍ͬͯɺq(z) ͰαϯϓϦϯ άͨ͜͠ͱʹΑΔ p(z) ͔ΒͷζϨΛิਖ਼͢ΔҼࢠͰ͋Δɻ
ͨͱ͑ɺ͋Δ l ʹରͯ͠ɺq(z(l)) ͕΄ͱΜͲ 1 ͩͬͨͱ͖ɺαϯϓϧ ͷதʹසൟʹ z(l) ͕ೖͬͯདྷΔ͜ͱ͕༧͞ΕΔɻ ͨͩ͠ɺp(z(l)) ͕খ͍͞߹ɺ͋·Γαϯϓϧ z(l) (11.19) ʹد༩ ͢Δ͖Ͱͳ͍ɻ ͜ͷͱ͖ɺrl = p(z(l))/q(z(l)) ͕খ͘͞ͳͬͯɺ͔ͨ͠ʹαϯϓϧ z(l) (11.19) ʹد༩͠ͳ͍ɻ 20 / 56
11.1.4 ॏαϯϓϦϯά ·ͨɺҎલʹ͋ͬͨ௨Γɺp(z) ͕ن֨ԽఆΛআ͍ͯΘ͔͍ͬͯΔͱ ͢Δɻ ͭ·Γɺp(z) = p(z)/Zp ͱͨ͠ͱ͖ʹɺp(z) Θ͔͍ͬͯΔ͕ɺZp
Θ͔͍ͬͯͳ͍ͱԾఆ͢Δɻ·ͨɺಉ༷ʹ q(z) = q(z)/Zq ͱ͢Δɻ ͜ͷͱ͖ɺ(11.19) ҎԼͷΑ͏ʹͳΔɻ E[f] = ∫ f(z)p(z) dz = ∫ f(z) p(z) q(z) q(z) dz ∼ 1 L L ∑ l=1 p(z(l)) q(z(l)) f(z(l)) = Zq Zp 1 L L ∑ l=1 p(z(l)) q(z(l)) f(z(l)) = Zq Zp 1 L L ∑ l=1 rl f(z(l)) (11.20) ͜͜Ͱɺ{z(l)} q(z) ͔ΒαϯϓϦϯά͞Εͨαϯϓϧͷू߹Ͱ͋Γɺ rl = p(z(l))/q(z(l)) ͱఆٛͨ͠ɻ 21 / 56
11.1.4 ॏαϯϓϦϯά ·ͨɺZp = ∫ p(z) dz ͱ Zq =
q(z)/q(z) Λ༻͍ΔͱɺZp /Zq ҎԼͷ Α͏ʹಉ͡αϯϓϧͷू߹ {z(l)} ΛͬͯɺۙࣅతʹܭࢉͰ͖Δɻ Zp Zq = ∫ 1 Zq p(z) dz = ∫ p(z) q(z) q(z) dz ∼ 1 L L ∑ l=1 rl (11.21) ΑͬͯɺE[f] ҎԼͷΑ͏ʹۙࣅͰ͖Δɻ E[f] ∼ Zq Zp 1 L L ∑ l=1 rl f(z(l)) ∼ L ∑ L m=1 rm · 1 L L ∑ l=1 rl f(z(l)) = L ∑ l=1 rl ∑ L m=1 rm f(z(l)) = L ∑ l=1 wl f(z(l)) (11.22) ͜͜Ͱɺwl ҎԼͰఆٛ͞ΕΔɻ wl = rl ∑ L m=1 rm (11.23) 22 / 56
11.1.4 ॏαϯϓϦϯά ͜ΕΑΓɺ͔֬ʹॏαϯϓϦϯάΛ༻͍Δͱɺظ E[f] Λۙࣅత ʹٻΊΔ͜ͱ͕Ͱ͖Δɻ ͨͩ͠ɺ͜ͷαϯϓϦϯάͰ༻͢ΔఏҊ q(z) αϯϓϧΛٻΊ ͍ͨ
p(z) ͱ͋Δఔࣅ͍ͯΔΛ༻͢Δඞཁ͕͋Δɻ ͨͱ͑ɺp(z) ͕͋Δαϯϓϧۭؒͷখ͞ͳൣғ A Ͱ 0 Ͱͳ͍Α͏ͳ Ͱ͋ΓɺఏҊ q(z) ͕ͦͷൣғ A Ͱ 0 Ͱ͋Δͱɺαϯϓϧͷू ߹ {z(l)} ͷதͰൣғ A ʹೖΔͷͳ͍ͷͰɺॏཁॏΈͷू߹ {rl } ͯ͢ 0 ͱͳΔɻ ͜ͷͱ͖ɺE[f] ͷۙࣅ 0 ͱͳͬͯ͠·͏ɻ ͞ΒʹɺE[f] ͕ 0 ͱͳΔͷ͕ਖ਼͍͔͠൱͔ (ਖ਼͍͠߹͋Δ) ΛଌΔ ࢦඪ͕ॏαϯϓϦϯάͰଘࡏ͠ͳ͍ͷ͕࠷ਂࠁͳͰ͋Δɻ 23 / 56
11.1.5 SIR غ٫αϯϓϦϯάͷɺ্͔Βԡ͑͞ΔΑ͏ͳ k ΛܾΊΔͱɺغ ٫͕ߴ͘ͳΔ͜ͱͩͬͨɻ ͜͜Ͱɺk ΛઃఆͤͣʹࡁΉํ๏Ͱ͋Δ SIR(Sampling-Importance-Resampling) ʹ͍ͭͯઆ໌͢Δɻ
·ͣɺఏҊ q(z) ͔Β L ݸͷαϯϓϧू߹ {z(l)} ΛಘΔɻ (Sampling) (11.23) ΑΓɺwl Λܭࢉ͢Δɻ(Importance) wl ∑ l wl = 1 Λຬͨ͢ͷͰɺࢄతͳ֬ͱΈͳͤΔͷͰɺͦ ͷ֬ pl = wl ʹै͏֬Ͱ {z(l)} ͔Β L ݸαϯϓϦϯά͢Δɻ (Resampling) ҎԼͰɺ͜ͷΑ͏ʹ࠶αϯϓϦϯά͞Εͨαϯϓϧू߹ {z(l)} L → ∞ ͰαϯϓϦϯά͍ͨ͠ p(z) ͔ΒͷαϯϓϦϯάʹͳ͍ͬͯ Δ͜ͱΛࣔ͢ɻ 24 / 56
11.1.5 SIR αϯϓϧۭ͕ؒҰมͷ߹Ͱɺ֬ pl = wl ͷྦྷੵ p(z ≤ a)
Λܭࢉͯ͠ΈΔɻ(֬ม z ͕ a ҎԼͱͳΔ֬) p(z ≤ a) (11.23) Λ༻͍Δͱ p(z ≤ a) = ∑ l:z(l)≤a wl = ∑ l I(z(l) ≤ a)p(z(l))/q(z(l)) ∑ m p(z(m))/q(z(m)) (11.25) ͱͳΔɻ ͜͜ͰɺI(z ≤ a) Ҿ͕ਅͷ࣌ 1 ͰɺͦΕҎ֎Ͱ 0 ͱͳΔؔͰ ͋Δɻ 25 / 56
11.1.5 SIR L → ∞ ͱ͢Δͱɺ{z(l)} ͱͱ q(z) ͔ΒαϯϓϦϯά͞Εͨαϯ ϓϧͳͷͰɺ
p(z ≤ a) = ∫ I(z ≤ a){p(z)/q(z)}q(z) dz ∫ {p(z)/q(z)}q(z) dz = ∫ I(z ≤ a)p(z) dz ∫ p(z) dz = ∫ I(z ≤ a)p(z) dz (11.26) ͱͳΓɺ͔֬ʹ L → ∞ ͰɺαϯϓϦϯά͍ͨ͠ p(z) ͷྦྷੵ ʹҰக͍ͯ͠Δɻ 26 / 56
11.1.6 αϯϓϦϯάͱ EM ΞϧΰϦζϜ EM ΞϧΰϦζϜͷ E εςοϓͷۙࣅܭࢉʹαϯϓϦϯά๏༻Ͱ ͖Δɻ EM
ΞϧΰϦζϜͷ E εςοϓͰɺӅΕม Z ͷࣄޙ p(Z|X, θold) ʹΑΔશσʔλର ln p(Z, X|θ) ͷظ Q(θ, θold) ΛٻΊΔɻ Q(θ, θold) = ∫ p(Z|X, θold) ln p(Z, X|θ) dZ (11.28) ͜ͷੵΛɺࣄޙ p(Z|X, θold) ͔Βͷαϯϓϧू߹ {Z(l)} Λͬ ͯɺҎԼͷ༗ݶͰۙࣅ͢Δɻ Q(θ, θold) ∼ 1 L L ∑ l=1 ln p(Z(l), X|θ) (11.29) ͦͯ͠ɺ͜ͷ Q(θ, θold) Λ༻͍ͯ M εςοϓΛ࣮ߦ͢Δɻ ͜ͷΑ͏ͳ EM ΞϧΰϦζϜΛϞϯςΧϧϩ EM ΞϧΰϦζϜͱ͍͏ɻ 27 / 56
11.2 Ϛϧίϑ࿈ϞϯςΧϧϩ ͜͜·Ͱղઆ͖ͯͨ͠αϯϓϦϯά๏Ͱɺαϯϓϧۭ͕ؒ࣍ݩͳΒ ༗ޮͰ͋Δ͕ɺߴ࣍ݩʹͳΔͱ্ख͍͔͘ͳ͍͜ͱ͕Θ͔͍ͬͯΔɻ ͜͜Ͱಋೖ͢ΔϚϧίϑ࿈ϞϯςΧϧϩ (MCMC) Ͱαϯϓϧۭؒ ͕ߴ࣍ݩͰ͋ͬͯΑ͘ػೳ͢Δɻ ɹ MCMC
ʹ͓͍ͯɺఏҊΛಋೖ͢Δɻ ͨͩ͠ɺ͜͜ͰͷఏҊݱࡏͷαϯϓϧ z(τ) ʹґଘ͠ɺq(z|z(τ)) ͱͳΓɺ࣍ͷαϯϓϧ z(τ+1) q(z|z(τ)) ͔ΒಘΒΕΔɻ(ޙ΄Ͳઆ໌͢ ΔϚϧίϑ࿈) ·ͨɺඪ p(z) ͔ΒαϯϓϧΛಘΔ͜ͱͰ͋Δ͕ɺ͜͜Ͱ p(z) = p(z)/Zp ͱ͠ɺp(z) Θ͔͍ͬͯΔ͕ɺZp ͷΘ͔Βͳ͍ͱ ͢Δɻ 28 / 56
11.2 Ϛϧίϑ࿈ϞϯςΧϧϩ ·ͣ MCMC ͷҰൠʹೖΔલʹɺ࠷؆୯ͳྫͰ͋Δ Metropolis Ξϧ ΰϦζϜΛղઆ͢Δɻ ఏҊͯ͢ͷαϯϓϧʹ͍ͭͯରশͰ͋Δͱ͢Δɻ ྫ͑
zA ͱ zB ʹରͯ͠ q(zA |zB ) = q(zB |zA ) (ରশͰͳ͍߹ 11.2.2 Ͱղઆ͢Δɻ) ·ͣɺαϯϓϧ z(τ) ͕༩͑ΒΕͯɺ࣍ͷαϯϓϧީิ z⋆ ͕ఏҊ q(z|z(τ)) ͔ΒಘΒΕͨͱ͢Δɻ ͦͷαϯϓϧީิ z⋆ ҎԼͷ֬ A(z⋆, z(τ)) Ͱडཧ͞ΕΔɻ A(z⋆, z(τ)) = min ( 1, p(z⋆) p(z(τ)) ) (11.33) ࣮Ͱɺ୯Ґ۠ؒ (0, 1) ͷҰ༷͔Βཚ u Λऔಘ͠ɺ A(z⋆, z(τ)) > u Ͱ͋Εɺαϯϓϧީิ z⋆ αϯϓϧͱͯ͠डཧ͞Εɺ ͦΕҎ֎ͷ߹غ٫͞ΕΔΑ͏ʹ͢Δɻ 29 / 56
11.2 Ϛϧίϑ࿈ϞϯςΧϧϩ ͜͜Ͱ (11.33) ΑΓɺp(z⋆) > p(z(τ)) Ͱ͋ΕɺA(z⋆, z(τ)) =
1 ͱͳΓɺ ͲΜͳཚ u Ͱඞͣ A(z⋆, z(τ)) > u ͱͳΔͷͰɺඞͣαϯϓϧީิ z⋆ αϯϓϧͱͯ͠डཧ͞ΕΔ͜ͱʹҙɻ ͠ɺαϯϓϧީิ z⋆ αϯϓϧͱͯ͠डཧ͞ΕͨΒɺz(τ+1) = z⋆ ͱ ͠ɺغ٫͞ΕͨΒ z(τ+1) = z(τ) ͱ͢Δɻ ͜ͷ͕ɺغ٫͞ΕͨΒ୯ʹαϯϓϧΛࣺͯΔغ٫αϯϓϦϯάͱͷҧ ͍Ͱ͋Δɻ (11.2.2 Ͱূ໌͢Δ͕ɺ) ҙͷ zA ͱ zB ʹରͯ͠ q(zA |zB ) ͕ਖ਼ͳΒ ɺz(τ) ͕ै͏ τ → ∞ ͰαϯϓϧΛಘ͍ͨ p(z) ʹۙͮ͘ɻ 30 / 56
11.2.1 Ϛϧίϑ࿈ MCMC ͷҰൠతͳٞΛ͢ΔͨΊɺ·ͣϚϧίϑ࿈ͷٞΛߦ͏ɻ ͦ͜Ͱɺ·ͣ m εςοϓͰͷ֬มΛ z(m) ͱ͢Δɻ ͦͯ͠ɺm
= 1, · · · , M ͱͨ͠ͱ͖ͷ֬มͷܥྻ z(1), · · · , z(M) ʹ ରͯ͠ɺҎԼͷੑ࣭ (ಠཱੑ) Λຬͨ͢ͱ͖ɺz(1), · · · , z(M) ΛϚϧίϑ ࿈ͱ͍͏ɻ p(z(m+1)|z(1), · · · , z(m)) = p(z(m+1)|z(m)) (11.37) ͭ·Γɺεςοϓ m + 1 ͷ֬աఔ͕Ұݸલͷεςοϓ m ΑΓલͷ εςοϓʹґଘ͠ͳ͍ͱ͍͏͜ͱͰ͋Δɻ ͜ͷ༷ࢠΛάϥϑͰද͢ͱɺҎԼͷΑ͏ʹͳΔɻ 31 / 56
11.2.1 Ϛϧίϑ࿈ m + 1 εςοϓͰͷαϯϓϧ z(m+1) Λൃੜͤ͞Δपล p(z(m+1))
p(z(m+1)) = ∑ z(m) p(z(m+1), z(m)) = ∑ z(m) p(z(m+1)|z(m))p(z(m)) (11.38) ͱͳΔɻ ͜͜ͰɺTm (z(m), z(m+1)) = p(z(m+1)|z(m)) ભҠ֬ͱ͍͏ɻ ಛʹભҠ͕֬εςοϓ m ʹґΒͳ͍ (Tm (z(m), z(m+1)) = T(z(m), z(m+1))) Ϛϧίϑ࿈ΛۉҰϚϧίϑ࿈ ͱ͍͍ɺࠓޙͦͷΑ͏ͳભҠ֬ʹݶఆͯ͠ΛਐΊΔɻ 32 / 56
11.2.1 Ϛϧίϑ࿈ Ҏ্ͷٞΑΓɺT(z(m), z(m+1)) ͕༩͑ΒΕΕɺϚϧίϑ࿈ʹै ͏֬ม࣍ͷΑ͏ʹͯ͠ൃੜͤ͞Δ͜ͱ͕Ͱ͖Δɻ 1. ॳظঢ়ଶ z(1) Λॳظ
(ྫ͑αϯϓϦϯά͕ՄೳͳఏҊ q(z)) ͔ΒαϯϓϦϯά͢Δ 2. m = 1, · · · , M − 1 ʹରͯ͠ɺભҠ֬ T(z(m), z(m+1)) ΑΓ z(m+1) Λൃੜͤ͞Δ ͜ͷΑ͏ʹͯ͠ൃੜͤͨ͞αϯϓϧ z(m+1) (11.38) ΑΓɺ͔֬ʹ p(z(m+1)) ͔Βൃੜͤͨ͞αϯϓϧͰ͋Δɻ 33 / 56
11.2.1 Ϛϧίϑ࿈ ͜ͷΑ͏ʹαϯϓϧ z(m) Λੜ͠ଓ͚ͯɺm → ∞ ͱͨ͠ͱ͖ʹ p(z(m)) ऩଋ͢Δ͔Ͳ͏͔͕ͱͳΔɻ
͜ΕϚϧίϑ࿈͕ΤϧΰʔυੑΛຬ͍ͨͯ͠Εɺऩଋ͢Δ͜ͱ͕ Θ͔͍ͬͯΔɻ Ϛϧίϑ࿈͕ΤϧΰʔυతͰ͋ΔͱɺنੑʢͲͷঢ়ଶ͔ΒͰ ҙͷঢ়ଶભҠͰ͖Δʣͱਖ਼࠶ؼੑʢҙͷঢ়ଶԿճͰભҠͰ͖ Δʣͱඇपظੑʢҙͷঢ়ଶҰճͷભҠͰݩʹΕΔʣΛશͯಉ࣌ʹ ຬͨ͢͜ͱΛݴ͏ɻ ͦͯ͠ΤϧΰʔυੑΛຬ͍ͨͯ͠ΔۉҰϚϧίϑ࿈Ͱɺm → ∞ ͱ ͨ͠ͱ͖ɺ p(z(m)) ҎԼͷৄࡉΓ߹͍݅Λຬͨ͢ p⋆(z) ʹऩଋ͢Δɻ p⋆(z)T(z, z′) = p⋆(z′)T(z′, z) (11.40) 34 / 56
11.2.1 Ϛϧίϑ࿈ Ҏ্Λ౿·͑ΔͱɺఏҊ q(z) ͔ΒɺϚϧίϑ࿈Λͬͯرͷ p⋆(z) ͔ΒͷαϯϓϧΛಘ͍ͨͳΒҎԼͷखॱͰαϯϓϦϯάΛߦ ͍͍͑ɻ 1.
(11.40) ͷৄࡉΓ߹͍݅ΛΈͨ͢Α͏ͳભҠ֬ T(z, z′) Λ༻ ҙ͢Δ 2. ॳظঢ়ଶ z(1) ΛఏҊ q(z) ͔ΒαϯϓϦϯά͢Δ 3. m = 1, · · · , M − 1 ʹରͯ͠ɺT(z(m), z(m+1)) ͔Β z(m+1) Λൃੜ ͤ͞Δ 4. m → ∞ ͷͱ͖ͷαϯϓϧ z(m) رͷ p⋆(z) ͔Βͷαϯϓ ϧͰ͋Δ 35 / 56
11.2.2 Metropolis-Hastings ΞϧΰϦζϜ ҎલɺMetropolis ΞϧΰϦζϜΛ MCMC ͷҰྫͱͯ͠հ͕ͨ͠ɺͦ Ε͕αϯϓϦϯά͍ͨ͠ p(z) ͔ΒͷαϯϓϦϯάʹͳ͍ͬͯΔ͜
ͱઆ໌͠ͳ͔ͬͨɻ ͜͜ͰɺMetropolis ΞϧΰϦζϜΛ֦ுͯ͠ɺఏҊ͕ରশͰͳ͍ ߹ (zA ͱ zB ʹରͯ͠ q(zA |zB ) ̸= q(zB |zA )) ͷΞϧΰϦζϜ (Metropolis-Hastings ΞϧΰϦζϜ) Λಋೖ͢Δɻ ·ͣɺαϯϓϧ z(τ) ͕༩͑ΒΕͯɺ࣍ͷαϯϓϧީิ z⋆ ͕ఏҊ qk (z|z(τ)) ͔ΒಘΒΕͨͱ͢Δɻ(ఴ͑ࣈ k ભҠઌ͕ෳ͋ͬͨ߹ ͷͨΊʹ͚͍ͭͯΔɻ11.3 Ͱ۩ମྫΛݟΔɻ) ͦͷαϯϓϧީิ z⋆ ҎԼͷ֬ Ak (z⋆, z(τ)) Ͱडཧ͞ΕΔɻ Ak (z⋆, z(τ)) = min ( 1, p(z⋆)qk (z(τ)|z⋆) p(z(τ))qk (z⋆|z(τ)) ) (11.44) 36 / 56
11.2.2 Metropolis-Hastings ΞϧΰϦζϜ ͜ͷ Metropolis-Hastings ΞϧΰϦζϜͰɺqk (z′|z) ʹΑΓαϯϓϧީ ิ͕બΕɺAk (z′,
z) ʹΑΓडཧ͞ΕΔ͔Ͳ͏͔ܾ·ΔͷͰɺભҠ֬ Tk (z, z′) = qk (z′|z)Ak (z′, z) ͱͳΔ͔Βɺ(11.44) ΑΓ p(z)Tk (z, z′) =p(z)qk (z′|z)Ak (z′, z) =p(z)qk (z′|z) · min ( 1, p(z′)qk (z|z′) p(z)qk (z′|z) ) =p(z)qk (z′|z) · min ( 1, p(z′)qk (z|z′) p(z)qk (z′|z) ) =min ( p(z)qk (z′|z), p(z′)qk (z|z′) ) =p(z′)qk (z|z′) · min ( p(z)qk (z′|z) p(z′)qk (z|z′) , 1 ) =p(z′)qk (z|z′)Ak (z, z′) = p(z′)Tk (z′, z) (11.45) ͱͳΔͷͰɺαϯϓϧΛಘ͍ͨ p(z) ͕ৄࡉΓ߹͍݅ (11.40) Λ ຬͨ͢ͷͰɺMetropolis-Hastings ΞϧΰϦζϜΛ܁Γฦ͢ͱɺp(z) ͔Β ͷαϯϓϧΛಘΔ͜ͱ͕Ͱ͖Δ͜ͱ͕Θ͔Δɻ 37 / 56
11.2.2 Metropolis-Hastings ΞϧΰϦζϜ ͜ͷ Metropolis-Hastings ΞϧΰϦζϜʹҙ͕͋Δɻ ྫ͑ɺҎԼͷਤͷΑ͏ʹɺରশͳఏҊ q(z) Λݱࡏͷঢ়ଶ z(τ)
Λ த৺ʹͨ͠ํΨεͱ͠ɺp(z) ํʹΑͬͯେ͖͘ҟͳΔภ ࠩΛ࣋ͭΨεͱ͢Δɻ(͕ p(z) Ͱɺ੨͕ q(z)) ఏҊͷεέʔϧ ρ ͕খ͍͞ͱغ٫Լ͕Δ͕ɺz ͕ z(τ) ͔Βେ͖ ͘มԽ͠ͳ͍ͨΊɺϥϯμϜΥʔΫΛͱΓɺܥྻ z(1), · · · ͷؒͰ͍ ૬ؔΛ࣋ͭɻ Ұํɺεέʔϧ ρ େ͖͗͘͢͠Δͱ p(z) ͕খ͍͞αϯϓϧΛऔͬͯ͘ ΔՄೳੑ͕ߴ͘ͳΓɺغ٫্͕͕Δɻ 38 / 56
11.3 ΪϒεαϯϓϦϯά ࣍ MCMC ͷҰྫͰ͋ΔΪϒεαϯϓϦϯάΛઆ໌͢Δɻ ޙͰड़ΔΑ͏ʹɺΪϒεαϯϓϦϯά Metropolis-Hastings Ξϧΰ ϦζϜͷಛघͳ߹Ͱ͋Δ͜ͱ͕Θ͔Δɻ αϯϓϦϯάΛ͍ͨ͠Λ
p(z) = p(z1 , · · · , zM ) ͱ͠ɺ֤εςοϓͰ z = (z1 , · · · , zM )T ͷதͷҰͭͷ zi Λ p(zi |z\i ) ͔ΒαϯϓϦϯ ά͠ɺߋ৽͢Δɻ ͜͜Ͱɺz\i z ͔Β zi ΛऔΓআ͍ͨͷͰ͋Δɻ 1. {zi : i = 1, · · · , M} ΛॳظԽ͢Δ 2. τ = 1, · · · , T ʹରͯ͠ҎԼΛߦ͏ɻ ▶ p(z1 |z(τ) 2 , · · · , z(τ) M ) ͔Β z(τ+1) 1 ΛαϯϓϦϯά ▶ p(z2 |z(τ+1) 1 , z(τ) 3 , · · · , z(τ) M ) ͔Β z(τ+1) 2 ΛαϯϓϦϯά . . . ▶ p(zM |z(τ+1) 1 , z(τ+1) 2 , · · · , z(τ+1) M−1 ) ͔Β z(τ+1) M ΛαϯϓϦϯά 39 / 56
11.3 ΪϒεαϯϓϦϯά ຊདྷαϯϓϧΛٻΊ͍ͨ p(z) = p(z1 , · · ·
, zM ) ͔ΒͷαϯϓϦϯά αϯϓϧۭ͕ؒߴ࣍ݩͰ͋ΔͨΊɺغ٫αϯϓϦϯάͳͲͰαϯϓ ϦϯάͰ͖ͳ͍ɻ ͔͠͠ɺ͖݅ p(zi |z(τ+1) 1 , · · · , z(τ+1) i−1 , z(τ) i+1 , · · · , z(τ) M ) αϯϓ ϧۭ͕ؒҰ࣍ݩͳͷͰɺغ٫αϯϓϦϯάͳͲͰαϯϓϧ z(τ+1) i Λಘ Δ͜ͱ͕Ͱ͖Δɻ ·ͨɺΪϒεαϯϓϦϯάͰͷ֤εςοϓͰɺzi ͷΈ͕มߋ͞ΕΔͷ ͰɺఏҊ qi (z∗|z) = p(z∗ i |z\i ) ͱͳΔɻ ͜͜Ͱɺz εςοϓલͷมͰ z∗ εςοϓલͷมͰ͋Γɺzi ͷ Έ͕มߋ͞ΕΔͷͰ z∗ \i = z\i Ͱ͋Δɻ 40 / 56
11.3 ΪϒεαϯϓϦϯά ·ͨɺ͜ͷఏҊ qk (z∗|z) = p(z∗ k |z\k )
ͱҰൠతͳੑ࣭ p(z) = p(zk |z\k )p(z\k ) Λ༻͍Δͱ p(z∗)qk (z|z∗) p(z)qk (z∗|z) = p(z∗ k |z∗ \k )p(z∗ \k )p(zk |z∗ \k ) p(zk |z\k )p(z\k )p(z∗ k |z\k ) = 1 (11.49) ͱͳΔɻ ͜͜Ͱɺz∗ \k = z\k Λ༻͍ͨɻ ͭ·Γɺ(11.44) ΑΓɺ Ak (z∗, z) = min ( 1, p(z∗)qk (z|z∗) p(z)qk (z∗|z) ) = min(1, 1) = 1 ͱͳΓɺΪϒεαϯϓϦϯά Metropolis-Hastings ΞϧΰϦζϜͷಛ घͳ߹ (ৗʹडཧ) Ͱ͋Δ͜ͱ͕Θ͔Δɻ ΑͬͯɺΪϒεαϯϓϦϯάΛଓ͚ΕɺαϯϓϧΛಘ͍ͨ p(z) ͔ΒͷαϯϓϧΛಘΔ͜ͱ͕Ͱ͖Δ͜ͱ͕Θ͔Δɻ 41 / 56
11.4 εϥΠεαϯϓϦϯά Metropolis ΞϧΰϦζϜͷεςοϓ෯Λখ͘͞औΓ͗͢Δͱϥϯ μϜΥʔΫΛى͜͠ɺେ͖͘औΓ͗͢Δͱغ٫্͕͕Δ͜ͱͰ ͋ͬͨɻ εϥΠεαϯϓϦϯάͰɺͷಛʹ߹Θͤͯεςοϓ෯Λࣗಈత ʹௐઅ͢Δ͜ͱ͕Ͱ͖Δɻ ͜͜ͰαϯϓϧΛಘ͍ͨͷܗਖ਼نԽఆΛআ͍ͯΘ͔͍ͬͯ Δͱ͢Δɻ(p(z)
ͷܗΘ͔͍ͬͯΔ) ·ͨɺαϯϓϧۭؒҰ࣍ݩͱ͢Δɻ 42 / 56
11.4 εϥΠεαϯϓϦϯά αϯϓϧऔಘͷखॱͱͯ͠ɺݱࡏͷΛ z ͱͨ͠ͱ͖ʹൣғ 0 ≤ u ≤ p(z)
͔ΒҰ༷ʹ u ΛαϯϓϦϯά͢Δɻ ͦͯ͠ u Λݻఆͯ͠ɺp(z) > u ͱͳΔΑ͏ͳ z ͷྖҬ͔Β࣍ͷ z Λந ग़͢Δɻ(p(z) = u ͰεϥΠε) 43 / 56
11.4 εϥΠεαϯϓϦϯά ͜ͷαϯϓϦϯάಉ࣌ ˆ p(z, u) ͔ΒαϯϓϦϯάΛߦ͏͜ͱͱ ͍͠ɻ ˆ p(z,
u) = { 1/Zp (0 ≤ u ≤ p(z)) 0 (otherwise) (11.51) ͨͩ͠ɺ Zp = ∫ p(z) dz Ͱ͋Δɻ ·ͨɺz ͷपล (11.51) ΑΓ ∫ ˆ p(z, u) du = ∫ p(z) 0 1 Zp du = p(z) Zp = p(z) ͱͳΔͷͰɺαϯϓϧ (u, z) Λ ˆ p(z, u) ͔Βಘͯɺu Λແࢹ͢Εر ͷ p(z) ͔Βͷαϯϓϧ z ͕ಘΒΕΔɻ 44 / 56
11.4 εϥΠεαϯϓϦϯά ͨͩ͠ɺ࣮ࡍ p(z) > u ͱͳΔΑ͏ͳ z ͷྖҬ͔Β࣍ͷ z
Λநग़͢ Δ͜ͱ͕ࠔͳ͜ͱ͕ଟ͍ɻ ͦ͜ͰɺԼͷਤͷΑ͏ʹݱࡏͷ z ͷΛ z(τ) ͱ͢Δͱɺͦͷ z(τ) ΛؚΉ ൣғ zmin ≤ z(τ) ≤ zmax ͷҰ༷͔Β z(τ+1) Λநग़͢Δํ๏͕͋Δɻ Ͱ͖Δ͚ͩ p(z) > u ͱͳΔΑ͏ͳ z Λ zmin ≤ z(τ) ≤ zmax ͷதʹؚΊ ΔΑ͏ʹൣғΛܾΊ͍͕ͨɺ͛͗͢Δͱغ٫্͕͕ͬͯ͠·͏ ͕͋Δɻ 45 / 56
11.4 εϥΠεαϯϓϦϯά ͦ͜ͰɺൣғͷܾΊํͱͯ͠ɺ·ͣ z(τ) ΛؚΉΑ͏ʹϥϯμϜʹൣғ w ΛܾΊΔɻ ͦͷ w ͕εϥΠε֎ʹग़ΔΑ͏ʹ֦ு͢Δɻ
ͦͷ֦ு͞Εͨൣғͷத͔Βαϯϓϧ z′ Λநग़͠ɺͦΕ͕εϥΠε ʹ͋Εαϯϓϧͱ͢Δɻ(z(τ+1) = z′) ͠εϥΠε֎ʹ͋Εɺൣғ w Λ (z(τ) ΛؚΉΑ͏ʹ)z′ Λͱ͢ ΔΑ͏ʹॖখ͢Δɻ ͦͯ͠৽ͨʹ z(τ+1) ͱͳΓ͏ΔީิΛൣғ w ͷத͔ΒऔΓग़͢ɻ 46 / 56
11.5 ϋΠϒϦουϞϯςΧϧϩΞϧΰϦζϜ ͜Ε·ͰͷٞͷதͰΘ͔ͬͨΑ͏ʹ Metropolis-Hastings ΞϧΰϦζ Ϝͷɺখ͞ͳεςοϓΛऔΔͱϥϯμϜΥʔΫΛҾ͖ى͜ ͠ɺεςοϓΛେ͖͘͢Δͱغ٫্͕͕Δ͜ͱͰ͋Δɻ ͜͜ͰɺཧγεςϜΛࢀߟʹ͠ɺغ٫Λখ͘͞อͬͨ··ɺε ςοϓΛେ͖͘ͱΔํ๏Λߟ͑Δɻ 47
/ 56
11.5.1 ྗֶܥ ·ͣɺϋϛϧτϯྗֶʹ͍ͭͯઆ໌͢Δɻ ͜Ε·Ͱࢄతͳεςοϓ τ Ͱͷ֬มΛ z(τ) ͱ͕ͨ͠ɺτ Λ࿈ଓ มʹ֦ு͠ɺ֬มΛ
τ ͷؔ z(τ) ͱ͢Δɻ ྗֶͰݴ͑ɺ͜ͷ τ ͕࣌ؒͰɺz(τ) ͕ମͷҐஔϕΫτϧͱͳΔɻ ·ͨɺz(τ) ͷ τ ඍ r = dz dτ (11.53) ӡಈྔ (ମͷͱಉͷྔ) Ͱ͋Δɻ ·ͨɺܥͷϙςϯγϟϧΤωϧΪʔΛ E(z) ͱ͢ΔͱɺମҎԼͷӡ ಈํఔࣜʹै͍ӡಈΛߦ͏͜ͱ͕ΒΕ͍ͯΔɻ dr dτ = − ∂E(z) ∂z (11.55) 48 / 56
11.5.1 ྗֶܥ ͜ΕΛϋϛϧτϯܗࣜͰ·ͱΊͳ͓͢ɻ ·ͣӡಈΤωϧΪʔ K(r) ΛҎԼͷΑ͏ʹఆٛ͢Δɻ K(r) = 1 2
∥r∥2 (11.56) ͞ΒʹɺϙςϯγϟϧΤωϧΪʔ E(z) ͱӡಈΤωϧΪʔ K(r) Λ͠ ͨશΤωϧΪʔΛҎԼͰఆٛ͢Δɻ H(z, r) = E(z) + K(r) (11.57) ͜͜ͰɺશΤωϧΪʔ H(z, r) ϋϛϧτϯؔͱ͍͏ɻ ͜ͷஈ֊ͰɺҐஔϕΫτϧ z ͱӡಈྔϕΫτϧ r ಠཱͳϕΫτϧͰ ͋Δ͜ͱʹҙɻ((11.53) ͱ͍͏ؔຬͨ͞ͳ͍ɻ) 49 / 56
11.5.1 ྗֶܥ ͜ͷϋϛϧτϯؔΛ༻͍Δͱɺܥͷํఔࣜ (11.53) ͱ (11.55) ҎԼ ͷ 2 ͭͷࣜͰදͤΔ͜ͱ͕Θ͔Δɻ
dz dτ = ∂H ∂r (11.58) dr dτ = − ∂H ∂z (11.59) ͞Βʹɺ(11.58) ͱ (11.59) Λຬͨ࣌͢ɺϋϛϧτϯؔͷ࣌ؒඍΛܭ ࢉ͢ΔͱҎԼͷΑ͏ʹ 0 ͱͳΔɻ dH(z, r) dτ = 0 (11.60) ͭ·Γɺ(11.58) ͱ (11.59) Λຬͨ͢ӡಈɺશΤωϧΪʔ H(z, r) Λอ ଘ͢Δɻ 50 / 56
11.5.1 ྗֶܥ ͞ΒʹɺԼͷਤͷΑ͏ͳ (z, r) ۭؒ (Ґ૬ۭؒ) ͷ͋ΔྖҬΛߟ͑Δɻ ͜ͷྖҬͷ֤͕ํఔࣜ (11.58)
ͱ (11.59) Λຬͨ͠ͳ͕ΒҠಈͨ͠ޙ ͷମੵɺݩͷମੵͱಉ͡Ͱ͋Δ͜ͱ͕ΒΕ͍ͯΔɻ(ϦϰΟϧͷ ఆཧ) 51 / 56
11.5.1 ྗֶܥ Ҏ্ͷϋϛϧτϯؔͷ࣌ؒෆมੑͱϦϰΟϧͷఆཧ͔ΒɺҎԼͷΑ ͏ʹఆٛ͞Εͨ֬Ґ૬্ۭؒͷ (11.58) ͱ (11.59) Λຬͨ͢ม Խʹରͯ͠ෆมͰ͋Δ͜ͱ͕Θ͔Δɻ p(z,
r) = 1 ZH exp (−H(z, r)) (11.63) ͜͜Ͱɺ֬มมޙن֨Խཱ͕͍ͯ͠ͳ͍ͱ͍͚ͳ͍ (۩ମతʹ PRML(1.27) ͷΑ͏ͳมΛ͢Δ) ͷͰɺϋϛϧτϯؔͷ ࣌ؒෆมੑ͚ͩͰ p(z, r) ෆมʹͳΒͳ͍͜ͱʹҙɻ (11.58) ͱ (11.59) Λຬͨ͢ (p(z, r) Λෆมʹ͢Δ)z ͱ r ͷతͳ࣌ ؒมԽ (ੵ) ͷ༩͑ํͱͯ͠ɺϦʔϓϑϩοάࢄԽͱ͍͏ํ๏ ͕͋Δɻ 52 / 56
11.5.1 ྗֶܥ ϦʔϓϑϩοάࢄԽͱɺҎԼͷΑ͏ͳߋ৽ํ๏Ͱ͋Δɻ ˆ r(τ + ϵ/2) =ˆ r(τ) −
ϵ 2 ∂E ∂z (ˆ z(τ)) (11.64) ˆ z(τ + ϵ) =ˆ z(τ) + ϵˆ r(τ + ϵ/2) (11.65) ˆ r(τ + ϵ) =ˆ r(τ + ϵ/2) − ϵ 2 ∂E ∂z (ˆ z(τ + ϵ)) (11.66) ͜ͷߋ৽ੵʹ͔͔ΘΒͣϦϰΟϧͷఆཧΛຬͨ͢ߋ৽ํ ๏Ͱ͋Δɻ͔͠͠ɺH ͷʹ ϵ ͷޡ͕ࠩग़Δɻ 53 / 56
11.5.2 ϋΠϒϦουϞϯςΧϧϩΞϧΰϦζϜ Ҏ্ͷϋϛϧτϯྗֶͱ Metropolis ΞϧΰϦζϜΛ༥߹ͤͨ͞ϋΠϒ ϦουϞϯςΧϧϩΞϧΰϦζϜΛಋग़͢Δɻ ·ͣɺॳظঢ়ଶΛ (z, r) ͱ͠ɺϦʔϓϑϩοάࢄԽͰߋ৽ͨ͠มΛ
(z⋆, r⋆) ͱ͠ɺҎԼͷ֬Ͱडཧ͢Δɻ min(1, exp {H(z, r) − H(z⋆, r⋆)}) (11.67) ͠ϋϛϧτϯྗֶΛશʹγϛϡϨʔτͨ͠ΒɺH ෆมͳͷͰඞ ͣडཧ͞ΕΔ͕ɺϦʔϓϑϩοάࢄԽΛ༻͍ͨΒɺੵʹΑΔޡ ࠩʹΑΓɺH ͕มԽ͢Δ߹͕͋ΔͷͰɺغ٫͞ΕΔ͜ͱ͕͋Δɻ ·ͨɺৄࡉΓ߹ཱ͍͕݅͢Δ͔Ͳ͏͔ԋश 11.17 Ͱٞͯ͠ ͍Δɻ 54 / 56
11.6 ؔͷਪఆ ຊষͷ΄ͱΜͲͷ߹ͰɺαϯϓϧΛٻΊ͍ͨن֨ԽఆΛআ͍ ͯΘ͔͍ͬͯΔͱԾఆͨ͠ɻ ͭ·ΓɺαϯϓϧΛٻΊ͍ͨ pE (z) Λ pE (z)
= 1 ZE exp (−E(z)) (11.71) ͱͨ࣌͠ʹ E(z) ͷؔܗΘ͔͍ͬͯΔ͕ɺZE ͷΘ͔Βͳ͍ͱ ͍͏͜ͱͰ͋Δɻ ͞Βʹɺ͜Ε·Ͱͷ͍ΖΜͳαϯϓϦϯά๏Ͱ ZE Θ͔Βͳͯ͘ ɺۙࣅతʹ pE (z) ͔ΒαϯϓϦϯάͰ͖͍ͯͨɻ ͨͩ͠ɺZE ͷ͕Θ͔ΔͱϞσϧͷൺֱ͕Ͱ͖ɺศརͳ͕࣌͋Δɻ ͜͜ͰɺZE ͷΛΔํ๏Λઆ໌͢Δɻ 55 / 56
11.6 ؔͷਪఆ ϞσϧൺֱͰΓ͍ͨͷɺ2 ͭͷϞσϧͷؔͷൺͰ͋Δɻ ͦ͜ͰɺҎԼͷ pG (z) Λߟ͑Δɻ pG (z)
= 1 ZG exp (−G(z)) ͜͜ͰɺpG (z) ͔ΒͷαϯϓϦϯάՄೳͰ͋ΔͱԾఆ͢Δɻ ͜ͷ࣌ɺൺ ZE /ZG pG (z) ͔Βͷαϯϓϧͷू߹ {z(l)} Λ༻͍ͯҎԼ ͷΑ͏ʹܭࢉͰ͖Δɻ ZE ZG = ∑ z exp (−E(z)) ∑ z exp (−G(z)) = ∑ z exp (−E(z) + G(z)) exp (−G(z)) ∑ z exp (−G(z)) = ∑ z exp (−E(z) + G(z))pG (z) =EpG [exp (−E + G)] ∼ 1 L ∑ l exp (−E(z(l)) + G(z(l))) (11.72) 56 / 56