Share
Embed
グラフ畳み込みネットワークを用いたNP困難問題に対する強化学習アプローチ / Training Graph Convolutional Networks by Reinforcement Learning for Solving NP-hard Problems
卒業論文の発表スライド
knshnb
February 13, 2019
Transcript
20 slides
-
άϥϑΈࠐΈωοτϫʔΫΛ༻͍ͨ NP ࠔʹର͢Δ ڧԽֶशΞϓϩʔν ࠤ౻ݚڀࣨ 4 Ѩ෦݈৴ 2019/02/13 1 / 20 -
എܠ ઌߦݚڀ ఏҊख๏ ·ͱΊ ༧උεϥΠυ 2 / 20 -
എܠ: NP ࠔ ▶ ݱ࣮తͳ࣌ؒͰ࠷దղΛٻ ΊΒΕͳ͍ͱ৴͡ΒΕͯ ͍Δ ▶ ex. φοϓαοΫ, ࠷େΫϦʔΫ ▶ NP શΛશͯؼணͰ͖ ΔͨΊɺ1 ͭͷ͕ղ͚Δ ͜ͱʹඇৗʹՁ͕͋Δ ▶ ܭࢉෳࡶੑཧ͚ͩͰͳ͘ɺ ͞·͟·ͳͰग़ݱ 0-1 φοϓαοΫ ֤ wi , hi ͕༩͑ΒΕͨ G = {1, 2, ..., N} ʹ͍ͭ ͯɺ max ∑ i∈G′ wi s.t. ∑ i∈G′ hi ≤ H ͱͳΔ G′ ⊂ G 3 / 20 -
NP ࠔͷΞϓϩʔν ▶ ਖ਼֬ͳΞϧΰϦζϜ ▶ FPTɺࢦͷఈΛ͑Δ etc. ▶ Ұൠʹࢦ࣌ؒ ▶ ਫ਼อূ͖ۙࣅΞϧΰϦζϜ ▶ ࣮༻తʹΑ͍ΞϧΰϦζϜ͕ݟ͔͍ͭͬͯͳ͍͕ ଟ͍ ▶ ώϡʔϦεςΟΫε ▶ ૣ͘ಈ࡞͠ɺൺֱతΑ͍ղ͕ٻΊΒΕΔ ▶ ݻ༗ͷࣝɺߴͳνϡʔχϯά͕ඞཁ ▶ ػցֶश ▶ ۙݚڀ͕ਐΜͰ͍Δ ▶ ݻ༗ͷࣝɺߴͳνϡʔχϯά͕ඞཁͳ͍ 4 / 20 -
NP ࠔͷΞϓϩʔν ▶ ਖ਼֬ͳΞϧΰϦζϜ ▶ FPTɺࢦͷఈΛ͑Δ etc. ▶ Ұൠʹࢦ࣌ؒ ▶ ਫ਼อূ͖ۙࣅΞϧΰϦζϜ ▶ ࣮༻తʹΑ͍ΞϧΰϦζϜ͕ݟ͔͍ͭͬͯͳ͍͕ ଟ͍ ▶ ώϡʔϦεςΟΫε ▶ ૣ͘ಈ࡞͠ɺൺֱతΑ͍ղ͕ٻΊΒΕΔ ▶ ݻ༗ͷࣝɺߴͳνϡʔχϯά͕ඞཁ ▶ ػցֶश ▶ ۙݚڀ͕ਐΜͰ͍Δ ▶ ݻ༗ͷࣝɺߴͳνϡʔχϯά͕ඞཁͳ͍ 5 / 20 -
എܠ ઌߦݚڀ ఏҊख๏ ·ͱΊ ༧උεϥΠυ 6 / 20 -
ઌߦݚڀ [Dai+ 2017] άϥϑ্ͷ NP ࠔΛڧԽֶशʹΑͬͯղ͘ϑϨʔϜ ϫʔΫ S2V-DQN ΛఏҊɻ ▶ ঢ়ଶͱͯ͠ɺάϥϑٴͼطʹબͨ͠ू߹Λ࣋ͬͯ ͓͘ ▶ structure2vec[Dai+ 2016] ͱ͍͏ख๏Ͱ֤ͷಛ ྔΛϕΫτϧԽ ▶ ͦΕͧΕͷΛબΜͩࡍʹಘΒΕΔߦಈՁؔΛɺ Q ֶशΛ༻͍ͯ܇࿅ ▶ ܇࿅ͨ͠ߦಈՁؔΛݩʹɺᩦཉʹղΛٻΊΔ 7 / 20 -
ઌߦݚڀ [Zhuwen+ 2018] ࠷େ҆ఆू߹Λڭࢣ͋ΓֶशʹΑͬͯղ͘ख๏ΛఏҊɻ ▶ ֤͕࠷దղʹؚ·ΕΔ֬Λग़ྗ͢Δؔ f (G; θ) ΛɺάϥϑΈࠐΈωοτϫʔΫͰϞσϧԽ ▶ (άϥϑ, ࠷దղ) ͱ͍͏܇࿅σʔλΛ༻͍ͨڭࢣ͋Γ ֶश ▶ ܇࿅σʔλɺSATLIB ͔Βؼணͤͨ͞ 1200 ͷά ϥϑΛ 40,000 ݸ༻ ▶ ܇࿅ͨ͠ωοτϫʔΫΛݩʹɺᩦཉ๏ͱ Guided Tree Search ͱ͍͏ಠࣗͷ୳ࡧΞϧΰϦζϜͰղΛٻΊΔ 8 / 20 -
άϥϑΈࠐΈωοτϫʔΫ [Kipf et al. 2016] ▶ άϥϑ্ͷࠐΈωοτϫʔΫͷ͏ͪͷ 1 ͭ ▶ ۙͷͷಛྔΛΈࠐΜͰ࣍ͷʹൖ͍ͯ͘͠ Figure: https://tkipf.github.io/graph-convolutional-networks/ ▶ ೖྗͱͯ͠αΠζͷҟͳΔҙͷάϥϑΛड͚औΕΔ 9 / 20 -
ઌߦݚڀ·ͱΊ S2V-DQN [Dai+ 2017] o ܇࿅σʔλͷੜ͕ ෆཁ o ϑϨʔϜϫʔΫͷҰൠ ੑ͕ߴ͍ x [Zhuwen+] ʹൺͯੑ ೳ͕Α͘ͳ͍ [Zhuwen+ 2018] o ಛघͳ܇࿅σʔλͰͷ ֶश͕༷ʑͳάϥϑʹ ରͯ͠ҰൠԽ o ੑೳ͕Α͍ x ܇࿅σʔλͷੜ͕ߴ ίετ ▶ NP ࠔͷ܇࿅ σʔλͷ༻ҙҰൠ తʹ؆୯Ͱͳ͍ =⇒ ੑೳͷྑ͍ڧԽֶशͷख๏ΛఏҊ 10 / 20 -
എܠ ઌߦݚڀ ఏҊख๏ ·ͱΊ ༧උεϥΠυ 11 / 20 -
࠷େ҆ఆू߹ ڧԽֶशͷઃఆΛߟ͍͑͢ NP ࠔͱͯ͠ɺ࠷େ҆ఆ ू߹ʹɻ Definition ҆ఆू߹: ҙͷ 2 ͕ޓ͍ʹྡ͍ͯ͠ͳ͍ू߹ ▶ ҎԼɺݟ͔ͭͬͨ҆ఆू߹ͷαΠζ͕େ͖͍΄Ͳྑ͍ͱ ධՁ͢Δ 12 / 20 -
ఏҊख๏ ▶ ղΛ 1 ͭ֬ఆͤ͞ɺࣗٴͼͦͷۙΛऔΓআ͍ͨ༠ಋ ෦άϥϑʹ͍ͭͯಉ͡Λ࠶ؼతʹߟ͑Δ ▶ άϥϑΈࠐΈωοτϫʔΫɺαΠζ͕ಈతʹมΘΔ άϥϑΛೖྗՄೳ =⇒ ֤εςοϓͰͦΕͧΕͷΛબͿ֬ f (G; θ) ɺάϥϑΈࠐΈωοτϫʔΫͰϞσϧԽͰ͖Δ ▶ ֶशख๏: REINFORCE 13 / 20 -
࣮ݧ ҎԼͷ 2 छྨͷֶशख๏ͰɺCitation Networks (Cora, Citeseer, Pumbed) ʹର͢ΔղΛٻΊͨɻ ▶ Cora ͷΈΛ༻ֶ͍ͯश ▶ n = 100, m = 200 ͷͨ͘͞ΜͷछྨͷϥϯμϜάϥϑΛ ༻ֶ͍ͯश Table: Citation Networks ͷαΠζ Name ล Cora 2708 5429 Citeseer 3327 4732 Pumbed 19717 44335 14 / 20 -
݁Ռ ▶ ϥϯμϜάϥϑΛ༻͍ͯ܇࿅ͨ͠Ϟσϧ͕ɺCora Λͬ ͯ܇࿅ͨ͠Ϟσϧͱಉ͡ύϑΥʔϚϯεΛݟͤͨ ▶ ڧԽֶशΛ༻͍ͨطଘख๏Λେ্͖͘ճΓɺڭࢣ͋Γֶ शΛ༻͍ͨ Zhuwen et al. ͷख๏ʹ͍ۙ݁Ռ͕ಘΒΕͨ Table: ൃݟ͞Εͨ҆ఆू߹ͷαΠζ طଘख๏ ఏҊख๏ άϥϑ S2V-DQN Zhuwen et al. ϥϯμϜ Cora Cora 1381 1451 1440 1440 Citeseer 1705 1867 1864 1864 Pubmed 15709 15912 15912 15912 15 / 20 -
എܠ ઌߦݚڀ ఏҊख๏ ·ͱΊ ༧උεϥΠυ 16 / 20 -
·ͱΊ ▶ ࠷େ҆ఆू߹ΛάϥϑΈࠐΈωοτϫʔΫͱڧԽ ֶशʹΑͬͯղ͘ख๏ΛఏҊͨ͠ ▶ ڭࢣ͋ΓֶशΛ༻͍ͨ࠷৽ͷख๏ʹഭΔ݁Ռ͕ಘΒΕͨ ▶ ༷ʑͳϥϯμϜάϥϑ্Ͱ܇࿅͢Δ͜ͱͰɺ࣮ੈքͷά ϥϑʹద༻Ͱ͖ΔϞσϧ͕ಘΒΕͨ ▶ ఏҊख๏ɺಉ͡Α͏ͳߏΛ࣋ͭଞͷάϥϑ্ͷ NP ࠔʹԠ༻Ͱ͖Δɻͦͷࡍɺڭࢣ͋ΓֶशΛ༻͍ ͨطଘख๏ͱൺΔͱ܇࿅σʔλΛ༻ҙ͢Δඞཁ͕ͳ͍ 17 / 20 -
എܠ ઌߦݚڀ ఏҊख๏ ·ͱΊ ༧උεϥΠυ 18 / 20 -
ֶशख๏ ▶ ༷ʑͳάϥϑ্ͰࢼߦΛ܁Γฦͯ͠ڧԽֶश ▶ ํࡦޯΞϧΰϦζϜͷ 1 ͭɺREINFORCE [Williams 1992] Λ༻͍ͯํࡦͷύϥϝʔλ θ Λ࠷దԽ ∇θ J(θ) = 1 N N ∑ n=1 T ∑ t=1 ∇θ log p(an t |sn t , θ)R(hn) ▶ ใु R(h) ɺಘΒΕͨ҆ఆू߹ͷαΠζΛਖ਼نԽͨ͠ ͷΛ༻͍Δ 19 / 20 -
ࠓޙͷ՝ ▶ branch-and-reduce ͱݺΕΔਖ਼֬ͳղΛٻΊΔΞϧΰϦ ζϜ [Akiba+ 2014] ͕ɺݱ࣮తͳάϥϑʹରͯ͠Ռ Λ͍ͯ͠Δ ▶ ্ͷख๏Ͱ͍͠άϥϑʹର࣮ͯ͠ݧ 20 / 20