Upgrade to Pro — share decks privately, control downloads, hide ads and more …

20201102RLArch.pdf

Sponsored · Your Podcast. Everywhere. Effortlessly. Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
Avatar for Tatsuya Matsushima Tatsuya Matsushima
November 02, 2020
180

 20201102RLArch.pdf

Avatar for Tatsuya Matsushima

Tatsuya Matsushima

November 02, 2020
Tweet

Transcript

  1. "DUJPOBOE1FSDFQUJPOBT%JWFSHFODF.JOJNJ[BUJPO • 4VCNJUUFEPO4FQ W MBTUSFWJTFE0DU W  • ஶऀਞ͕݁ߏ͍͢͝ʢ(PPHMF#SBJO 6$-

    %FFQ.JOE 5SPOUPʣ • IUUQTBSYJWPSHBCT • *$-3VOEFSSFWJFXIUUQTPQFOSFWJFXOFUGPSVN JE+C"RTGC:T+Z ॻࢽ৘ใ  1MB/FU%SFBNFSʢम࢜Ϙε͕'SJTUPOʣ ࣗ༝ΤωϧΪʔݪཧʢ'&1ʣ -BZFS/PSN"EBNʜ %FFQ.JOEͰ͍Ζ͍Ζ
  2. ೖྗͷ֬཰ม਺ ɼͦΕҎ֎ͷʢજࡏʣม਺  ύϥϝʔλ  • ࠷దԽͷաఔͰ௚઀มߋ͢Δ͜ͱ͕Ͱ͖Δ΋ͷશͯ • ྫʣදݱʹؔ͢Δे෼౷ܭྔɼϞσϧͷύϥϝʔλɼํࡦͷύϥϝʔλ ࣮ࡍͷ෼෍

    ʢBDUVBMEJTUSJCVUJPOʣ • ੜ੒աఔ Λදݱ • ࣮ࡍͷ෼෍͸ະ஌ͷ৔߹΋͋Δɹྫʣσʔλ෼෍ • ύϥϝʔλ ͷӨڹΛड͚ͯมԽ͢Δ৔߹΋͋Δ λʔήοτͷ෼෍  • γεςϜͷؔ͢Δ޷·͍͠ઃఆʢQSFGFSFODFʣΛදݱ • ʢͪΌΜͱॻ͔Εͯͳ͍͕ʣͬͪ͜΋ύϥϝʔλ ͷӨڹΛड͚Δ x z ϕ pϕ x, z ∼ pϕ (x, z) ϕ τ(x, z) ϕ ֬཰ม਺΍෼෍ͷදه 
  3. ύϥϝʔλ ΛมԽͤͯ͞ɼશͯͷม਺ʹؔ͢Δ ࣮ࡍͷ෼෍ ΛͰ͖Δ͚ͩλʔήοτ ʹ͚ۙͮΔ໰୊  ͱͯ͠ఆࣜԽ • SFWFSTF,-΍JOGPSNBUJPOQSPKFDUJPOͱ΋ݺ͹ΕΔ •

    ʢ༨ஊʣSFWFSTF,-WTGPSXBSE,-ͷ࿩΋େࣄ • ͸Ϟσϧʢ͚ͩʣͰ͸ͳ͘ੜ੒աఔΛදݱͯ͠Δ͜ͱʹ஫ҙ • Ұ෦͸ਅͷ෼෍Ͱະ஌ɼҰ෦͸ύϥϝʔλ ʹΑΓมԽ͠͏Δ • ྫʣදݱֶशͷ৔߹ɼ ͸σʔλͱ৴೦ͷಉ࣌෼෍ɼ ͸જࡏม਺Ϟσϧ • ྫʣ੍ޚͷ৔߹ɼ ͸ݱࡏͷํࡦͰͷيಓͷ෼෍ɼ ͸يಓͷVUJMJUZ ϕ pϕ τ minϕ KL [pϕ (x, z)∥τ(x, z)] pϕ ϕ pϕ τ pϕ τ +PJOU,-.JOJNJ[BUJPO 
  4. +PJOU,-Λ෼ղ  • ୈ߲ɿજࡏม਺ͷ৴೦ Λ޷·͍͠෼෍ ʹ͚ۙͮΔ੍໿ • ୈ߲ɿ૬ޓ৘ใྔ ͷม෼Լք •

     • ૬ޓ৘ใྔͷԼքͷ࠷খԽ㱺૬ޓ৘ใྔͷ࠷େԽ • ॊೈͳ ͷ଒Λ࢖͏΄ͲλΠτͳԼք KL [pϕ (x, z)∥τ(x, z)] KPJOUEJWFSHFODF = EKL [pϕ (z ∣ x)∥τ(z)] SFBMJ[JOHMBUFOUQSFGFSFODFT − E [ln τ(x ∣ z) − ln pϕ (x)] JOGPSNBUJPOCPVOE pϕ (z|x) τ(z) I[x; z] I[x; z] ≥ I[x; z] − EKL [p(x ∣ z)∥τϕ (x ∣ z)] = E [ln τϕ (x ∣ z) − ln p(x)] τϕ (x|z) -BUFOU1SFGFSFODF  *OGPSNBUJPO#PVOET
  5. ΤʔδΣϯτ͕ೖྗ ΛมߋͰ͖ͳ͍৔߹ɼ MBUFOUQSFGFSFODFͷ ͷ߲͸ύϥϝʔλԽ͞Εͳ͍ʢఆ਺ʣ • ྫʣݻఆͷσʔληοτ͔Βͷֶश ͜ͷͱ͖ɼઌ΄Ͳͷ໨తؔ਺͸ม෼ਪ࿦ʹ͓͚Δࣗ༝ΤωϧΪʔ ʢ&-#0ʣʹରԠ  •

    ͰೖྗΛ࠶ߏ੒͠ͳ͕Βɼ Λࣄલ෼෍ ʹ͚ۙͮΔ x pϕ (x) min ϕ {EKL [pϕ (z ∣ x)∥τ(z)] − E [ln τ(x ∣ z)]} τ(x|z) pϕ (z|x) τ(z) 7BSJBUJPOBM'SFF&OFSHZ  *OGPSNBUJPO#PVOET
  6. +PJOU,-ͷଞͷ෼ղ  • ୈ߲ɿೖྗͷ৚݅෇͖෼෍ Λ޷·͍͠෼෍ ʹ͚ۙͮΔ • ୈ߲ɿ૬ޓ৘ใྔ ͷม෼Լք •

    ࠓճ͸જࡏۭؒͰදݱʢ Ͱ৚͚͖݅ͮͯͬ͞ͱಉ༷ͳٞ࿦Ͱٻ·Δʣ • ೖྗΛ؍ଌͨ͠ޙͷ৴೦ ͱ؍ଌ͢Δલͷ৴೦ ͷࠩʢ৘ใήΠϯʣΛදݱ • Լք͸ ͕࣮ࡍͷ৚݅෇͖෼෍ ΛΑۙ͘ࣅ͍ͯ͠Δ΄ͲλΠτ 㱺ม਺ͷؔ܎Λֶश͢Δ΄Ͳ૬ޓ৘ใྔΛ࠷େԽʢදݱྗ๛͔ͳ Λ࢖͏ͱ༗རʣ ૬ޓ৘ใྔͷ࠷େԽ͸ɼೖྗͷྑ͍જࡏදݱͷֶशͱɼજࡏදݱʹͱͬͯ༗ӹͳೖྗͷ ୳ࡧͷ྆ํΛઆ໌ʂ KL [pϕ (x, z)∥τ(x, z)] KPJOUEJWFSHFODF = EKL [pϕ (x ∣ z)∥τ(x)] SFBMJ[JOHJOQVUQSFGFSFODFT − E [ln τ(z ∣ x) − ln pϕ (z)] JOGPSNBUJPOCPVOE pϕ (x|z) τ(x) I[x; z] z τ(z|x) pϕ (z) τ(z|x) p(z|x) τ(z|x) *OQVU1SFGFSFODF  *OGPSNBUJPO#PVOET
  7. ͸ΤʔδΣϯτͷ޷·͍͠ೖྗͷ෼෍Λදݱ͢Δ΋ͷ • ͭ·ΓΤʔδΣϯτ͸ ʹऩଋͤ͞Α͏ͱ͢Δ • %JWFSHFODFNJOJNJ[BUJPOʹΑͬͯΤʔδΣϯτ͸ࣗ਎ͷϞσϧΛ࢖ͬ ͯ༧ଌͰ͖ΔΑ͏ͳೖྗΛݟ͚ͭΑ͏ͱ͢Δ • ࣗ਎ͷϞσϧͷ଒Ͱ͸༧ଌͰ͖ͳ͍ೖྗΛආ͚Α͏ͱ͢Δ •

    ใु ʹରͯ͠ ͱ͢Ε͹λεΫͷใुΛಋೖՄೳ • ใुͱͯ͠पล໬౓Λଧͪফ߲͢໨Λ༩͑ͯɼೖྗͷ޷·͠͞ͷ߲Λ औΓআ͘͜ͱ΋Մೳ • ͜ͷลΓ͸Ͱઆ໌ τ(x) τ(x) r(x) exp (r(x)) *OQVUQSFGFSFODF  .PEFMBT1SFGFSFODF
  8. %JWFSHFODFNJOJNJ[BUJPOͷදݱֶश΍୳ࡧʹؔ࿈͢Δ૬ޓ৘ใྔͷԼք͸දݱྗ๛͔ͳ λʔήοτΛ༻͍Δ΄ͲλΠτʢʣ • Ϟσϧ͕༩͑ΒΕͨݩͰͷೖྗͷपล໬౓ͷNBUDIJOH • ͜ͷͱ͖ɼΤʔδΣϯτ͸ೖྗΛे෼ʹ༧ଌͰ͖ͯɼ ֎ཚʹରԠͰ͖Δࣗવͳฏߧঢ়ଶʢFDPMPHJDBMOJDIFʣʹऩଋ ࣅͨΑ͏ͳ֓೦ͱͯ͠TVSQSJTFNJOJNJ[BUJPO͕͋Δ • Ϟσϧ͕༩͑ΒΕͨݩͰͷೖྗͷपล໬౓ͷ࠷େԽ

    • ͜ͷͱ͖ɼOJDIFTFFLJOH<'SJTUPO >ͱݺ͹ΕΔ • ֎ཚ͕ͳ͍ͱҰ୴୳ࡧͨ͜͠ͱͷ͋Δࣗ໌ͳղʹؕΔ • ʢদౢղऍɿͨͿΜ͜Ε͸ʮ҉͍෦԰໰୊ʯͱݴΘΕͯΔ΋ͷʣ %JWFSHFODFNJOJNJ[BUJPOͩͱɼೖྗͷपล෼෍Λ߹ΘͤΑ͏ͱ͢ΔͨΊɼ ༧ଌͰ͖ΔશͯͷೖྗΛΧόʔ͢ΔΑ͏ͳ޿͍OJDIFʹͳΔΑ͏ʹଅ͞ΕΔ • දݱྗ๛͔ͳੈքϞσϧΛ༻͍Δͱ༧ଌՄೳੑ͕ߴ·Γ޿͍OJDIFʹʢλεΫใु͸௥Ճతʣ /JDIF4FFLJOH  .PEFMBT1SFGFSFODF
  9. աڈͷೖྗ Λ؍ଌ͠ɼকདྷͷೖྗ Λ ؍ଌ͢ΔલͷKPJOU,-ͷ࠷దԽΛߟ͑Δ • ͷ૬ޓ৘ใྔͷԼքΛࣜมܗ x< x> E [ln

    τ(z ∣ x) − ln pϕ (z)] JOGPSNBUJPOCPVOE = E [ln τ(z ∣ x) − ln pϕ (z ∣ x<) + ln pϕ (z ∣ x<) − ln pϕ (z)] ≥ E [ln τ(z ∣ x) − ln pϕ (z ∣ x<)] GVUVSFJOGPSNBUJPOCPVOE + E [ln τ (z ∣ x<) − ln pϕ (z)] QBTUJOGPSNBUJPOCPVOE ࣌ܥྻ΁ͷ֦ு 
  10. ૬ޓ৘ใྔͷԼքͷࣜมܗΛར༻ͯ͠ ࠷খԽ͍ͨ͠KPJOU,- ͷ্ք͕ٻ·Δ  • +PJOU,-͸ɼաڈͷೖྗͷ༗༻ͳදݱͷֶशͱকདྷͷ༗༻ͳೖྗͷ ୳ࡧͷ྆ํΛಉ࣌ʹදݱͰ͖͍ͯΔ KL [pϕ (x,

    z)∥τ(x, z)] KL [pϕ (x, z)∥τ(x, z)] ≤ EKL [pϕ (z ∣ x<) ∥τ(z)] SFBMJ[JOHQBTUMBUFOUQSFGFSFODFT − E [ln τ (x< ∣ z) − ln pϕ (x<)] SFQSFTFOUBUJPOMFBSOJOH + EKL [pϕ (x> ∣ x< , z) ∥τ (x> ∣ x<)] SFBMJ[JOHGVUVSFJOQVUQSFGFSFODFT − E [ln τ(z ∣ x) − ln pϕ (z ∣ x<)] FYQMPSBUJPO ࣌ܥྻ΁ͷ֦ு 
  11. ʢิ଍ʣSFWJTJPO W Ͱ௥Ճ͞Εͨઆ໌ ʹ͸Ұ౓΋؍ଌ͞Εͳ͍΋ͷͱɼ࣌ؒͷܦաͱͱ΋ʹ ؍ଌ͞ΕΔ΋ͷ͕͋Δ • Ұ౓΋؍ଌ͞Εͳ͍ɿજࡏঢ়ଶͷਪఆ஋΍Ϟσϧύϥϝʔλʣ • ࣌ؒܦաͱͱ΋ʹ؍ଌɿ֬཰తͳߦಈ΍εΩϧ •

    ͜ΕΒͷม਺Λબ୒͢Δ͜ͱ͸ɼຊ౰͸ҼՌతͳհೖʹͳΔͷͰ ͸ɼະ؍ଌͷજࡏม਺ ͱɼ؍ଌ͞Εͨજࡏม਺ Ͱ৚͚݅ͮ ΒΕͨ Ͱ෼ׂ͢΂͖͕ͩɼ؆୯ͷͨΊʹলུ͍ͯ͠Δ z z z> EP(z< ) z> 
  12. ʢิ଍ʣSFWJTJPO W Ͱ௥Ճ͞Εͨઆ໌ ʹ͸Ұ౓΋؍ଌ͞Εͳ͍΋ͷͱɼ࣌ؒͷܦաͱͱ΋ʹ ؍ଌ͞ΕΔ΋ͷ͕͋Δ • Ұ౓΋؍ଌ͞Εͳ͍ɿજࡏঢ়ଶͷਪఆ஋΍Ϟσϧύϥϝʔλʣ • ࣌ؒܦաͱͱ΋ʹ؍ଌɿ֬཰తͳߦಈ΍εΩϧ •

    ͜ΕΒͷม਺Λબ୒͢Δ͜ͱ͸ɼຊ౰͸ҼՌతͳհೖʹͳΔͷͰ ͸ɼະ؍ଌͷજࡏม਺ ͱɼ؍ଌ͞Εͨજࡏม਺ Ͱ৚͚݅ͮ ΒΕͨ Ͱ෼ׂ͢΂͖͕ͩɼ؆୯ͷͨΊʹলུ͍ͯ͠Δ z z z> EP(z< ) z> 
  13. "DUVBMɿ  5BSHFUɿ  ྫʣ#//Λ༻͍ͨڭࢣ͋Γֶश • ೖྗɿը૾ ɼΫϥε  •

    σʔληοτͷHMPCBMͳදݱͱͯ͠જࡏతͳύϥϝʔλ Λਪ࿦ pϕ (x, y, w) = pϕ (w) ⏟ CFMJFG ∏ i p (xi , yi) EBUB τ(x, y, w) ∝ τ(w) ⏟ QSJPS ∏ i τ (yi ∣ xi , w) MJLFMJIPPE x = {xi } y = {yi } w ม෼ਪ࿦ʢ7BSJBUJPOBM*OGFSFODFʣ 
  14. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • ໨తؔ਺͕ࣗ༝ΤωϧΪʔʢ&-#0ʣʹͳΔ Λσϧλ෼෍ ͱ͢Δͱ  • ͭ·Γ࠷໬ਪఆʹͳΔʢλʔήοτΛ ͱͨ͠ͱ͖ͱ౳Ձʣ

    KL [pϕ ∥τ] = KL [pϕ (w)∥τ(w)] DPNQMFYJUZ − E[ln τ(y ∣ x, w)] BDDVSBDZ + E[ln p(x, y)] DPOTUBOU pϕ (w) pϕ (w) → δϕ (w) ≐ {1JGw = ϕFMTF0} KL [pϕ (w)∥τ(w)] DPNQMFYJUZ − E[ln τ(y ∣ x, w)] BDDVSBDZ → ln τ(ϕ) − E[ln τ(y ∣ x, ϕ)] DPNQMFYJUZ ≐ E [−ln τϕ (y ∣ x)] QBSBNFUFSJ[FEUBSHFU τϕ (y|x) ม෼ਪ࿦ʢ7BSJBUJPOBM*OGFSFODFʣ 
  15. "DUVBMɿ  5BSHFUɿ  ྫʣ7"& • ೖྗɿը૾  • ରԠ͢Δજࡏม਺

    Λਪ࿦ pϕ (x, z) = ∏ i p (xi) ⏟ EBUB pϕ (zi ∣ xi) FODPEFS τϕ (x, z) = ∏ i τϕ (xi |zi) EFDPEFS τ (zi) ⏟ QSJPS x = {xi } z = {zi } ঈ٫ਪ࿦ʢ"NPSUJ[FE*OGFSFODFʣ 
  16. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • ໨తؔ਺͕ࣗ༝ΤωϧΪʔʢ&-#0ʣʹͳΔ • ༧ଌͷ߲ʢୈ߲ʣ͕૬ޓ৘ใྔͷԼք • දݱ͕ೖྗʹؔͯ͠ҙຯΛ࣋ͭΑ͏ʹଅ͢ KL [pϕ

    ∥τϕ] = E [ln τϕ (x ∣ z) − ln p(x)] JOGPSNBUJPOCPVOEʢদౢෛ߸͕͍ΔͷͰ͸ʁʣ + EKL [pϕ (z ∣ x)∥τ(z)] DPNQMFYJUZ ঈ٫ਪ࿦ʢ"NPSUJ[FE*OGFSFODFʣ 
  17. "DUVBMɿ  5BSHFUɿ  ྫʣ)..Ͱকདྷͷೖྗ ͕ܽଛ͍ͯ͠Δ৔߹ • ೖྗ  •

    ରԠ͢Δજࡏม਺ Λਪ࿦ pϕ (x, z) = p (x> , x<) EBUB pϕ (z ∣ x<) CFMJFG τϕ (x, z) = τϕ (x< ∣ z) MJLFMJIPPE τϕ (x> ∣ z) QSFEJDUJPO τ(z) ⏟ QSJPS x> x = {x< , x> } z = {zi } কདྷͷೖྗʢ'VUVSF*OQVUTʣ 
  18. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • ͱͯ͠কདྷͷೖྗ ͕࠷దԽʹ ؔΘΒͳ͍Α͏ʹ͍ͯ͠ΔʢVODPOUSPMMFEGVUVSFʣ KL [pϕ ∥τϕ] =

    KL [pϕ (x< , z) ∥τϕ (x< , z)] WBSJBUJPOBMJOGFSFODF + EKL [p (x> ∣ x<) ∥τϕ (x> ∣ z)] VODPOUSPMFEGVUVSF τϕ (x> ∣ x< , z) = τϕ (x> ∣ x<) x> ঈ٫ਪ࿦ʢ"NPSUJ[FE*OGFSFODFʣ 
  19. "DUVBMɿ  5BSHFUɿ  ྫʣܾఆ࿦తͳ੍ޚΛ͢Δ৔߹ • ೖྗͷܥྻ ͸ΤʔδΣϯτ͕ύϥϝʔλ ΛมԽͤͯ͞ ෦෼తʹม͑ΒΕΔྫʣܾఆ࿦తͳํࡦɼ੍ޚଇɼϓϥϯʜ

    pϕ (x) = ∏ t pϕ (xt ∣ x1:t−1) DPOUSPMMFEEZOBNJDT τ(x) = ∏ t τ (xt ∣ x1:t−1) QSFGFSFODFT x = {xt } ϕ ʢܾఆ࿦తͳʣ੍ޚ 
  20. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • QSFGFSFODFͷର਺ͱೖྗͷΤϯτϩϐʔͷ࠷େԽ • $G .BY&OU3-  ͸ߦಈͷΤϯτϩϐʔ࠷େԽ •

    1SFGFSFODFͷઃఆ • ɿ,-DPOUSPM<,BQQFO > • ɿ,-SFHVMBSJ[FEDPOUSPM<5PEPSPW > • ɿFYQFDUFESFXBSE<4VUUPO > • ͜Ε͸DVSJPTJUZͷ߲͕ফ͑ΔΑ͏ͳઃఆʢ·ΕͳೖྗΛ໌ࣔతʹEJTDPVSBHFͯ͠୳ࡧΛ੍ݶʣ • ೖྗͷΤϯτϩϐʔͷਪఆɿٖΧ΢ϯτɼ4LFX'JUɼ3/%ɼ*$.ʜ KL [pϕ ∥τ] = − ∑ t E [ln τ (xt ∣ x1:t−1)] FYQFDUFEQSFGFSFODFT + H [pϕ (xt ∣ x1:t−1)] DVSJPTJUZ τ (xt ∣ x1:t−1) ∝ exp (r (xt)) τ (xt ∣ x1:t−1) ∝ exp (r (xt)) τ′ (xt ∣ x1:t−1) τ (xt ∣ x1:t−1) ∝ exp (r (xt)) pϕ (xt |x1:t−1 ) ʢܾఆ࿦తͳʣ੍ޚ 
  21. "DUVBMɿ  5BSHFUɿ  ྫʣ֬཰తͳ੍ޚΛ͢Δ৔߹ • NBY&OU3-ɾ3-BTJOGFSFODFͱ΋͍͏ • ೖྗͷܥྻ ɼߦಈͷܥྻ

     • աڈͷೖྗߦಈ ͕؍ଌ͞Ε͍ͯΔઃఆ pϕ (x, a) = ∏ t p (xt ∣ x1:t−1 , a1:t−1) FOWJSPONFOU pϕ (at ∣ x1:t , a1:t−1) QPMJDZ τ(x, a) ∝ ∏ t exp (r (xt)) SFXBSE p (xt ∣ x1:t−1 , a1:t−1) FOWJSPONFOU τ (at) ⏟ BDUJPOQSJPS x = {xt } a = {at } ax< .BY&OUSPQZ3- .BY&OU3- 
  22. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • ߦಈۭؒͰͷ੍໿ͱظ଴ใुͷ߲ʹͳΔ • λʔήοτʹ؀ڥͷμΠφϛΫε͕ೖΔͷͰͷDVSJPTJUZͷ߲͸ଧͪ ফ͞ΕΔ • ߦಈͷࣄલ෼෍ 

    • Ұ༷෼෍ɿ42-ɼ4"$ • աڈͷೖྗ΍ҎલͷλεΫͷ஌ࣝΛ౷߹ɿ%JTUSBM KL [pϕ ∥τ] = ∑ t EKL [pϕ (at ∣ x1:t , a1:t−1) ∥τ (at)] DPNQMFYJUZ − E [r (xt)] FYQFDUFESFXBSE τ(a) .BY&OUSPQZ3- .BY&OU3- 
  23. "DUVBMɿ  5BSHFUɿ  ྫʣ֬཰తͳ੍ޚʹ͓͚ΔHFOFSBMJ[FEFNQPXFSNFOU • աڈͷೖྗߦಈ ͕؍ଌ͞Ε͍ͯΔઃఆ • ೖྗ͔ΒߦಈΛ༧ଌ͢Δλʔήοτͷ෼෍ʢSFWFSTFQSFEJDUPSʣ

    pϕ (x, a) = ∏ t p (xt ∣ x1:t−1 , a1:t−1) FOWJSPONFOU pϕ (at ∣ x1:t , a1:t−1) QPMJDZ τ(x, a) ∝ ∏ t τϕ (at ∣ x1:T , a1:t−1) SFWFSTFQSFEJDUPS ax< &NQPXFSNFOU 
  24. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • (FOFSBMJTFEFNQPXFSNFOU͸૬ޓ৘ใྔ ͷԼք KL [pϕ ∥τϕ] = EKL[p(x

    ∣ a)∥τ(x)] DPOUSPM − E [ln τϕ (a ∣ x) − ln pϕ (a)] HFOFSBMJ[FEFNQPXFSNFOU E [ln τϕ (a ∣ x) − ln pϕ (a)] HFOFSBMJ[FEFNQPXFSNFOU ≥ ∑ t E[ln τϕ (at ∣ x, a1:t−1) EFDPEFS − ln pϕ (at ∣ a1:t−1) TPVSDF ] I[a; x] = E [ln pϕ (a|x) − ln pϕ (a)] &NQPXFSNFOU 
  25. "DUVBMɿ  5BSHFUɿ  ྫʣ࣌ؒతʹந৅Խ͞ΕͨTLJMMͰ৚͚݅ͮΒΕͨํࡦΛֶश • ೖྗͷܥྻ ɼߦಈͷܥྻ ɼεΩϧͷܥྻ 

    • εΩϧ͸ݻఆͷ෼෍ʢ͔աڈͷೖྗͷؔ਺ʣ͔Βબ୒ • ߦಈ͸աڈͷೖྗͱݱࡏͷεΩϧʹैͬͯબ୒ pϕ (x, a, z) ≐ T/K ∏ k=1 pϕ (zk) BCTUSBDUQPMJDZ T ∏ t=1 pϕ (at ∣ x1:t , a1:t−1 , z⌊t/K⌋) QPMJDZ p (xt ∣ x1:t−1 , a1:t−1) FOWJSPONFOU τϕ (x, a, z) ∝ T/K ∏ k=1 τϕ (zk ∣ x) SFWFSTFQSFEJDUPS T ∏ t=1 τ (at) ⏟ BDUJPOQSJPS x = {xt } a = {at } z = {zt } εΩϧͷൃݟʢ4LJMM%JTDPWFSZʣ 
  26. +PJOU,-ʹͦΕͧΕ୅ೖ͢Δͱ  • 4LJMMEJTDPWFSZ͸εΩϧ ͱೖྗ ͷ૬ޓ৘ใྔ ͷԼք • ΤʔδΣϯτ͸ଟ༷ͳೖྗܥྻΛ࣮ݱ͠ͳ͕ΒɼҟͳΔεΩϧͰܥ ྻ͕ॏͳΒͳ͍Α͏ʹεΩϧΛֶश͢Δ

    • ྫʣ%*:"/<&ZTFOCBDI >ɼ%"%4<4IBSNB > KL [pϕ ∥τϕ] = EKL[p(x ∣ a)∥τ(x)] DPOUSPM + EKL [pϕ (a ∣ x, z)∥τ(a)] DPNQMFYJUZ − E [ln τϕ (z ∣ x) − ln pϕ (z)] TLJMMEJTDPWFSZ z x I[z; x] εΩϧͷൃݟʢ4LJMM%JTDPWFSZʣ 
  27. "DUVBMɿ  5BSHFUɿ  ྫʣ&YQFDUFEJOGPSNBUJPOHBJOΛ࠷େԽ͢ΔߦಈΛબ୒ • ֎తͳใु͕ͳ͍৔߹ʹੈքΛޮ཰తʹֶश͢ΔͨΊ • ೖྗɿ 

    • HMPCBMͳදݱΛ࣋ͭજࡏతͳύϥϝʔλɿ  pϕ (x, w) = pϕ (w) ⏟ CFMJFG ∏ t pϕ (xt ∣ x1:t−1) DPOUSPMMFEEZOBNJDT τ(x, w) = τ(w) ⏟ QSJPS ∏ t τ (xt ∣ x1:t−1 , w) MJLFMJIPPE x = {xi } w *OGPSNBUJPO(BJO 
  28. +PJOU,-ʹͦΕͧΕ୅ೖ্ͯ͠քΛऔΔ   • ͱ͸ҟͳΓɼະདྷͷೖྗʹӨڹͰ͖Δ • աڈͷೖྗʹ͍ͭͯ༗ӹͳදݱΛֶश͠ɼະདྷͷදݱʹؔͯ͠༗ӹʹͳΔΑ͏ͳೖྗΛ୳ࡧ͢Δ • ϞσϧύϥϝʔλʹΑͬͯΤʔδΣϯτ͕؀ڥͷϧʔϧΛݟ͚ͭΔΑ͏ʹ͢Δٞ࿦ •

    BDUJWFJOGFSFODF<'SJTUPO >ɼ7*.&<)PVUIPPGU >ɼ."9<4IZBN ^ɼ1MBO&YQMPSF<4FLFS > KL [pϕ ∥τϕ] ≤ EKL [pϕ (w ∣ x<) ∥τ(w)] TJNQMJDJUZ − E [ln τϕ (x< ∣ w) − ln pϕ (x<)] SFQSFTFOUBUJPOMFBSOJOH +EKL [pϕ (x> ∣ x< , w) ∥τϕ (x> ∣ x<)] DPOUSPM − E [ln τϕ (w ∣ x) − ln pϕ (w ∣ x<)] JOGPSNBUJPOHBJO E [ln τϕ (w ∣ x) − ln pϕ (w ∣ x<)] JOGPSNBUJPOHBJO ≥ ∑ t′ >t E [ln τϕ (w ∣ x1:t′ ) − ln pϕ (w ∣ x1:t′ −1)] JOUSJOTJDSFXBSE *OGPSNBUJPO(BJO 
  29.