210122_サンプリングの高速化へ向けた4次元変分法の発展と応用

249 Views

November 21, 23

スライド概要

profile-image

研究してるひと https://www.eri.u-tokyo.ac.jp/people/ito/

シェア

またはPlayer版

埋め込む »CMSなどでJSが使えない場合

関連スライド

各ページのテキスト
1.

サンプリングの高速化へ向けた4次元変分法の発展と応用 伊藤 伸一 東京大学 地震研究所 東京大学大学院 情報理工学系研究科 スマートサンプリング講演会 2021.01.22 © 伊藤伸一 01/40

2.

自己紹介 学歴・職歴 2015.03 大阪大学 大学院理学研究科 博士(理学) ・破壊のパターン形成の数理 2015.04—2018.08 東京大学 地震研究所 特任研究員 2018.08—現在 東京大学 地震研究所 助教 2018.12—現在 東京大学 大学院情報理工学系研究科 助教(兼務) ・大規模4次元変分法データ同化の数理・応用 ・金属材料の粒成長パターン ・断層摩擦パラメータの空間パターン 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 02/40

3.
[beta]
データ同化
モデル
<latexit sha1_base64="9Cy5FDqhVcuvIh7OEYFm+wU107A=">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</latexit>

事後分布

データ

p(xt | D)

D = {D1 , D2 , ...}

予測の高精度化、

<latexit sha1_base64="9Cy5FDqhVcuvIh7OEYFm+wU107A=">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</latexit>

p(xt | D)

<latexit sha1_base64="uqBeDhpsi6lxGlvwkHHdVQpc2sc=">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</latexit>

実験・観測デザインの

逆問題、…

最適化

事後分布(に比例した量)を1点評価するのに1回のモデルの時間発展計算が必要
・小さいモデルを使う?

→ 表現能力の低下

・評価回数を極力減らす?

→ 事後分布の精度低下

→ モデルの規模・目的に適したアルゴリズムの選択が大切

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

04/40

4.
[beta]
本日の話題
アルゴリズムの話題を中心にお話しします。
・4次元変分法データ同化
・サンプリング法との融合可能性
相補的な高速化へ向けた試み(模索中)
休憩
<latexit sha1_base64="GpdAVcHh9JDzrelSbAp+4vMUz30=">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</latexit>

・大規模不確実性評価

p(✓)

Ito et al., Phys. Rev. E, 94, 043307 (2016)
Ito et al., STAM,18:1, 857-869 (2017)

✓ˆ
<latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit>

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

05/40

5.

4次元変分法データ同化と サンプリング法との融合可能性 © 伊藤伸一 06/40

6.

データ同化アルゴリズムあれこれ 逐次データ同化 <latexit sha1_base64="9Cy5FDqhVcuvIh7OEYFm+wU107A=">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</latexit> 非逐次データ同化 p(xt | D) <latexit sha1_base64="YZ41l0LDB4d7yMkbgRl7814ENRI=">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</latexit> <latexit sha1_base64="9Cy5FDqhVcuvIh7OEYFm+wU107A=">AAACu3ichVFNLwNRFD3Gd30VG4lNo6nUpnlFQiQSwcKySpFQzcx4eMxXZl6bVtM/4A+QWGliIX6GjYWthZ8gliQ2Fu5MJxGacicz777z7rn3vDmaYwhPMvbcprR3dHZ19/RG+voHBoeiwyNbnl10dZ7TbcN2dzTV44aweE4KafAdx+WqqRl8Wztd8c+3S9z1hG1tyorD86Z6ZIlDoauSoH0nWS5UZW3PFAex1alCNM5SLIhYc5IOkzjCyNjRR+zhADZ0FGGCw4Kk3IAKj55dpMHgEJZHlTCXMhGcc9QQIW6RqjhVqISe0veIdrshatHe7+kFbJ2mGPS6xIwhwZ7YLXtjD+yOvbDPlr2qQQ9fS4VWrcHlTmHofGzj41+WSavE8TfrT80Sh5gPtArS7gSIfwu9wS+dXbxtLGQT1UlWZ6+k/5o9s3u6gVV612/WefbqDz0aafH/WKJlhUSZ5tuBAx5qZGX6t3HNydZ0Kj2Tml6fjS8th6b2YBwTSJJzc1jCGjLI0RwXl7hGXVlUdOVEMRqlSlvIGcWPUIpfmGCeig==</latexit> p(x0 | D) p(xt | D) ・概形と時間発展を評価 事後分布をサンプル群の時間発展で評価 ・小中規模モデル向き たくさんのシミュレーションを 同時に走らせる必要がある(並列化可) ・EnKF, 粒子フィルタ など 2021.01.22 スマートサンプリング講演会 x0 <latexit sha1_base64="Jx+cW+SJxE+Pa7o4z0mLl2CNdgI=">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</latexit> ・概形も時間発展も評価しない 初期値の事後分布のMAP解だけを探索 ・大規模モデル向き シミュレーション1つに 計算資源を集中できるため ・4次元変分法 © 伊藤伸一 07/40

7.

大規模モデル 偏微分方程式 <latexit sha1_base64="5Eie73Ys07UcLXeUEy82VPz67Ac=">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</latexit> @y = r (cry) @t @c =0 @t (y1 , c1 ) (y2 , c2 ) 常微分方程式 <latexit sha1_base64="6O3s4p4FnPINseuVsA+v55PCATM=">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</latexit> ... <latexit sha1_base64="Vz5tsOaASbTxiZ8j+uThnvYzxSg=">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</latexit> <latexit sha1_base64="wWKI6lX86S2Dj26rEurav0pQBN4=">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</latexit> L <latexit sha1_base64="m7GGtaDFdh43+pFgjcmmwRRVns8=">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</latexit> 空間方向 離散化 <latexit sha1_base64="okOfZccAOeFGBRquAtq2PDLqkhU=">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</latexit> dx = F (x) dt x = (y1 , y2 , . . . , c1 , c2 , . . . ) > <latexit sha1_base64="CMdOeZoBuiIpVP6uog8dmfb70BA=">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</latexit> ↑ 格子点上の物理量を全て並べたベクトル L 2 L 3 <latexit sha1_base64="+DuWf/G/Q1qN3rxQKEZpH5U5+jE=">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</latexit> <latexit sha1_base64="thjWTd6RihTzFbYsp8xNrm6WOqA=">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</latexit> 空間1方向に100格子点とると, 2次元問題 100x100 格子点 = O(10,000) 変数 ! 3次元問題 100x100x100 格子点 = O(1,000,000) 変数 !! 大規模モデルの事後分布を 効率よく評価するアルゴリズム <latexit sha1_base64="Hj6tScwPBCrBnunp6KYUmhnhFwA=">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</latexit> 出典:Google Earth 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 08/40

8.

4次元変分法 モデルの初期値に関する事後分布から事後確率最大解(MAP解)を抽出する手法 Dデータと + !t t = h(xt )の関係 Dt = h(xt ) + !t シミュレーションモデル <latexit sha1_base64="yXpqNkT0w6ffWorbjXV9yQ9bbtU=">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</latexit> <latexit sha1_base64="yXpqNkT0w6ffWorbjXV9yQ9bbtU=">AAACzHichVE9S8NQFD3G789WXQQRgqWiCOVVBUUQRB2cRK3VQpWSxGcN5ovktaihk+DgH3BwUnAQN/0JLg6uDv0J4qjg4uBNGhAV9YXknXvePfeel6s6hu4Jxqp1Un1DY1NzS2tbe0dnVyze3bPu2SVX41nNNmw3pyoeN3SLZ4UuDJ5zXK6YqsE31L354HyjzF1Pt601ceDwLVMpWvqOrimCqEJ8YKHgi4o8I+8O7wdoRB6VN22TF5UgKsQTLMXCJf8E6QgkEK1lO/6ATWzDhoYSTHBYEIQNKPDoySMNBoe4LfjEuYT08JyjgjbSliiLU4ZC7B59ixTlI9aiOKjphWqNuhj0uqSUkWSP7Iq9sHt2zZ7Y+6+1/LBG4OWAdrWm5U4hdtKXeftXZdIusPup+tOzwA6mQq86eXdCJriFVtOXD09fMtOrSX+IXbBn8n/OquyObmCVX7XLFb569ocflbwEfyz5a4bAPvW3wwl4CEaZ/j64n2B9LJUeT42tTCRm56KhtqAfgximyU1iFotYRpb6HOEKN7iVliQh+VKllirVRZpefFnS8QeM6KR2</latexit> 初期値 x0 = ✓ h:観測演算子 事後分布 p(✓ | D) / p(✓) <latexit sha1_base64="oWi9GfktOIkXj88PGSyzQpA7glc=">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</latexit> p(✓):事前分布 <latexit sha1_base64="WGN2PBPUYctwd9zIQsrpAM3Un4c=">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</latexit> ノイズ <latexit sha1_base64="9gygmdn+QQW5JH9b9JNHqm5Ff9o=">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</latexit> <latexit sha1_base64="3tnKdZhOroykKE9AkYQLgeiHt+E=">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</latexit> <latexit sha1_base64="BVEjgQf0iX35/BBzir1DM3XQaXQ=">AAACyXichVE9LwRRFD3G9+cuGqHZ2KzQbN4iISqhkWh8LRIrm5nxrBdvPsy83WCziUTnDyhUJCSi4D9oFFqFnyBKEo3CndlJBMGdzLz7zrvn3vPmGK4UvmLssUarratvaGxqbmlta++IxTu7ln2n6Jk8azrS8VYN3edS2DyrhJJ81fW4bhmSrxjb08H5Sol7vnDsJbXn8nVLL9hiU5i6Iigf7805Fi/o+bKqJHK+sBI7gzmjKCVXQ/l4kqVZGImfSSZKkohizonfI4cNODBRhAUOG4pyCR0+PWvIgMElbB1lwjzKRHjOUUELcYtUxalCJ3SbvgXarUWoTfugpx+yTZoi6fWImUCKPbBL9sLu2BV7Yu+/9iqHPQIte7QaVS5387GjnsW3f1kWrQpbn6w/NStsYjzUKki7GyLBLcwqv7R//LI4sZAqD7Az9kz6T9kju6Ub2KVX83yeL5z8occgLcEfS/1aobBL853QAR8VsjLz3bifyfJwOjOSHp4fTU5ORaY2oQ/9GCTnxjCJGcwhS3MOcIFr3Giz2o62q+1XS7WaiNONL6EdfgAiXKPw</latexit> Y q(Dt !t ⇠ q(•) h(xt )) t2T obs MAP解 T obs:観測時刻のセット <latexit sha1_base64="SMmXtD5Kzifn6YlK+hRImb/o8g8=">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</latexit> <latexit sha1_base64="e075dcW3ikGHqSQBkD9Kp8T8K60=">AAAOqniclZc9b9NAGMeflrcSIG1hQWKxSFs1Eo0uBUHF1JKqare+N7QuVexcEqtO7NpOyIvCB0DsDEwgMSA2BhbYWPgCDP0IiLFILAw8vpgkTS7JE1e1z889v///ucs5OWu2abgeY6cjoxcuXrp8Zexq6Nr1G+Hxicmbu65VdHS+o1um5SS1lMtNo8B3PMMzedJ2eCqvmXxPO074/Xsl7riGVdj2KjY/zKeyBSNj6CkPQ0cTM/as6uW4l1LzRlpZjiqq7Vi2Zyn8WW0uEfRF60cTERZj4lC6G/GgEYHgWLcmL38GFdJggQ5FyAOHAnjYNiEFLv4dQBwY2Bg7hBrGHGwZop9DHULIFjGLY0YKo8d4zuLdQRAt4L2v6QpaRxcT/x0kFZhmP9gHdsa+s4/sJ/vbU6smNPxaKnjVGiy3j8Zf3t76M5BS8JzCmnMtrm/VHmRgQVRrYPW2iPjj0Bt8qfr6bOvx5nRthr1jv3AEb9kp+4ZjKJR+6+83+OYbVJ/uWZMHZdS3xBy75yryaypgznMxS3lRdQE/lxrG/VnPikgZGT/yf0wWOvj3DkaUIO9FM1PF2TbwzgjcKB7+qGkejcxuj5BYTxzHqUIJrzrMQgRXUTRQ1VCxLNqaWA/pppMCUyJzCnXqHTplEl2WshUSW5GyVRJblbJFEluUsg6JdaSsRmI1KWuTWFvKnpDYEynrklhXyuZIbE7KrpLYVSmbILEJKbtMYpel7AqJXZGyayR2Tcruk9h9KZsksUkpu0Ril3o8vxyfEoukwKQKlvgFy2IeRUNty5evcyP4LqeptfJlan4exzNVrZUvUzPFb6aGEapeO9Fr9unVVfvUVsZ5oKn4mTKFYSrpXUee+K2uikyZQlbMGX0FtPLlo7LFGjHFroI6unZGprrdtq4aOhlk1bZ4N5MmuaelflttT0W7Xyteb+4teu1g/P2JJXrrzftZvOZwrI19mL8Di+H+YgFbecwy+vZGB+yYfH0Pe6wBfqzNT+bIhnTM4ZV3jLJbd67DeVA/xdvfuRbF7p3uXOrR/3CIMWfQpTTwc21pdnsO6+g/HeW+fnMDHOXzG8L3s3jn21h3Y3c+Fr8fm994EFl8ErypjcEduIt1xOERLOJ+ZB120OkVfIIv8DV8L7wZfho+aKSOjgTMLTh3hNP/ANbl42M=</latexit> コスト関数 C(✓) = <latexit sha1_base64="dGQ7FytXyMRO7Mw9F/722V4PPcI=">AAAO03iclZfPTxNBFMcfqIioBfRi4mUjYNpDyRSNEhMTEELgJL8KBBdJt0zbDdvdZXdbF5p6MJ5MPHvwpIkH41n/AS/+Ax6IVz0Yj5h48eCb6dqWdkofS9idffM+3++b6Ww7a7iW6QeMHfb0njl7ru98/4WBi5cuxwaHhq+s+U7Jy/J01rEcb8PI+NwybZ4OzMDiG67HM0XD4uvG7ozoXy9zzzcdezXYd/lWMZO3zZyZzQQY2h56OBPXgwIPMgntvpbULSevufVIUtP9UnG7Euimra0+rugBD4OKY/jValWTuXvxWeyuJgvxUFwTie2hETbO5KG1N1JRYwSiY9EZ7vsEOuyAA1koQRE42BBg24IM+Pj3CFLAwMXYFlQw5mHLlP0cqjCAbAmzOGZkMLqL5zzePYqiNt4LTV/SWXSx8N9DUoMx9pW9Z0fsC/vAfrK/HbUqUkPUso9Xo8Zyd3vwxbWVP10pDc8ZrLnQ4E6sOoAcTMpqTazelRExjmyNLx+8Olq5tzxWucnesl84gjfskH3GMdjl39l3S3z5NaqPdawpgBD1HTnH/rGKRE025jyRs1SUVdv4uVQwLmY9LyMhMiLyf0wOOoh7DyNalPe0nqnjbJt4Z0ZuFA8xappHLbPdY0CuJ47j1KGM1yzEYQRXUSJSNVAxlG1DroedupMGozJzFHWqLTohiQ6V7D6J3VeyByT2QMmWSGxJyXok1lOyBok1lKxLYl0lu0di95SsT2J9JVsgsQUlO09i55XsDImdUbKzJHZWyc6R2Dklu0BiF5TsJondVLIbJHZDyU6T2OkOzy/Hp8QhKTClgiN/wfKYR9HQm/LV69yMvstpao18lZrI43imqjXyVWqW/M00MELVayY6zT69uoMTagtxHmgqIlOlcJpKOtdRJH6r6zJTpZCXc0ZfAY189ahcuUYsuaugjq6ZUamuNq2rmk4OWb0p3s7skNx3lH4rTU9Fs18jXq3vLTrtYMT+xJG91fp9HK8FHGttHyZ2YOO4v5jEVhGzzBN7E112TEI/wB6nix9r8lM5slM6FvDKW0bZrptsce7WT/EWO9eS3L3Tncsd+u+cYsw5dCl3/Vwbmu2ep3UUT0d4ol+yi6N6fgfw/SzV+jbW3libGE/dGp9Yuj0y9SB6U+uH63AD60jBXZjC/cgipNHpI3yD7/Ajlo5VYs9iz2upvT0RcxWOHbGX/wAQ4/Ne</latexit> log p(✓) p(✓ | D) / e X C(✓) log q(Dt h(xt )) t2T obs 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 09/40

9.

4次元変分法による初期値更新 x0 = ✓ C(✓) = <latexit sha1_base64="3tnKdZhOroykKE9AkYQLgeiHt+E=">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</latexit> X t2T obs <latexit sha1_base64="ZS3Z0hDusosxrFqeLGFiyaBbqyI=">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</latexit> :データ ✓guess <latexit sha1_base64="YnLegehdSrKrSmzbu3/KDfP7jUU=">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</latexit> Time 適当に選んだ x0 <latexit sha1_base64="AqZRKhAp+ail3X3ypq29XDB6qtc=">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</latexit> = ✓guess ではデータと全く合わないが,,, 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 10/40

10.

4次元変分法による初期値更新 x0 = ✓ C(✓) = <latexit sha1_base64="3tnKdZhOroykKE9AkYQLgeiHt+E=">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</latexit> X t2T obs <latexit sha1_base64="ZS3Z0hDusosxrFqeLGFiyaBbqyI=">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</latexit> :データ ✓ˆ <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> ✓guess <latexit sha1_base64="YnLegehdSrKrSmzbu3/KDfP7jUU=">AAAOmXiclZfNThNRFMcP+IVVC2hMTNg0FgwsaG7RKHEFQgi4AspHA0OaTnvbTpjODDPTOm1TH8AXcOEKExfK3hdw4wu44BGMS0zcuPDM7dgO7S09HcLMnXPP7/8/9/ZOe0e1dM1xGTsfGb12/cbNW2O3I3fu3ouOT0ze33PMip3juzlTN+20mnW4rhl819Vcnactm2fLqs731eMVv3+/ym1HM40dt2bxo3K2aGgFLZd1MZSZeKi4Je5mMw3F5Z7bKFa44zSbmYk4SzBxxHobyaARh+DYNCdvfgUF8mBCDipQBg4GuNjWIQsO/h1CEhhYGDuCBsZsbGmin0MTIshWMItjRhajx3gu4t1hEDXw3td0BJ1DFx3/bSRjMMN+sM/sgn1nZ+wn+9tXqyE0/FpqeFVbLLcy4+8epf4MpGJ4zmLNpQ53ZdUuFGBRVKth9ZaI+OPItfhq/f1F6uX2TOMJ+8h+4QhO2Tn7hmMwqr9zn7b49gdUn+lbkwse6ptijp1LFfk1GZjzRsxSWVRt4OfSwLg/60UR8ZDxI//HZKKDf29jJBbkvW1nKjjbGt5pgRvFwx81zaOV2esREeuJ4zgVqOI1B7MQx1U0F6iqqOiJtirWQ77tFINpkTmNOs0uHY9Ee1K2RmJrUrZOYutStkJiK1LWJrG2lFVJrCplLRJrSdkTEnsiZR0S60jZEoktSdl1ErsuZVdI7IqUXSWxq1J2jcSuSdkNErshZQ9I7IGUTZPYtJRdJrHLfZ5fjk+JSVJgUgVT/IIVMY+ioYTy5etcC77LaWqdfJman8fxTFXr5MvUdPGbqWKEqhcm+s0+vbr6FbV5OA80FT9TpjBMJf3rKBO/1RWRKVMoijmjr4BOvnxUllgjuthVUEcXZmSqO6F11dIpIKuE4r1MnuSel/qlQk9F2K8Tb7b3Fv12MP7+xBS9zfb9LF5LONbWPszfgSVwf7GIrTJmaVf2zg3YMfn6LvaYA/xYyE/myIZ0LOGVd42yV3e+y3lQP8Xb37lWxO6d7lzt0/98iDEX0KU68HPtaPZ6DuvoPx3elX7zAxzl8xvB97Nk99tYb2NvIZF8mljYehZfehW8qY3BFDzGOpLwApZwP7IJu+jUgFP4AmfRqehydD36upU6OhIwD+DSEU39A+IK3mc=</latexit> Time 勾配法による最適化により データに適合する初期値 X C(✓) の勾配 = <latexit sha1_base64="ZS3Z0hDusosxrFqeLGFiyaBbqyI=">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</latexit> 2021.01.22 スマートサンプリング講演会 x0 = ✓ˆ を探す <latexit sha1_base64="Lo0LEFFzTOuOLCRfu+MtakxyflQ=">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</latexit> の計算が必要 t2T obs © 伊藤伸一 11/40

11.

Adjoint法による勾配計算 勾配法によって最適化したいが,,, C(✓) = log p(✓) <latexit sha1_base64="dGQ7FytXyMRO7Mw9F/722V4PPcI=">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</latexit> Adjoint法 L=C+ <latexit sha1_base64="78IbWdMIbrr+VecwGQKtTmPH9ds=">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</latexit> Z T X logq(D q(Dtt h(x h(xtt)) )) log t2T obs ✓ で直接微分ができない ✓ ◆ d T :終端時刻 > dt t f (xt ) xt t:Adjoint変数 dt <latexit sha1_base64="JhwuSzHapToMJpcRjI3puCA2jKQ=">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</latexit> <latexit sha1_base64="jRRHIl4aQaDCLy6WjaPD1Ye14kE=">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</latexit> <latexit sha1_base64="ajy51y/F7329LyhzgkGTAYF0l28=">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</latexit> 0 <latexit sha1_base64="MTTozS6+KSrcznvpKqdNQgdB5ko=">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</latexit> (ラグランジュの未定乗数) 変分をとって係数比較 Adjoint モデル d dt <latexit sha1_base64="1kqcoTy2iTY3iSE4F+wj0lrXbFY=">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</latexit> t = (r x f ) > t + X t0 2T obs (t 0 t )r xt0 C <latexit sha1_base64="h5FlaEbyb5cnhAtKyCqna4vs1GU=">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</latexit> T =0 t=Tから時間後ろ向きに解くことで、目的の勾配が得られる 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 12/40

12.
[beta]
時間後ろ向きに計算する
Adjoint モデル

d
dt

t

= (r x f )>

t +

X

t0 2T obs

<latexit sha1_base64="1kqcoTy2iTY3iSE4F+wj0lrXbFY=">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</latexit>

(t

t0 )r xt0 C
<latexit sha1_base64="h5FlaEbyb5cnhAtKyCqna4vs1GU=">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</latexit>

T

=0

t
<latexit sha1_base64="RWXuDuJHXJvlN8xJDO9R8AqtrUs=">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</latexit>
<latexit

Input

✓
<latexit sha1_base64="jRRHIl4aQaDCLy6WjaPD1Ye14kE=">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</latexit>

Output

<latexit sha1_base64="pbpm7LokrN6XiD4C/gJ6Clb8cQ0=">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</latexit>

dC
d✓

t
<latexit sha1_base64="RWXuDuJHXJvlN8xJDO9R8AqtrUs=">AAAKkXiclZbPTxNBFMcfiIrrD0AuJl4aC8ZTMzX+qqeihEC8UEqhgSVkdzttN2x3l91thTb4B+hV48GTJh6Mf4MnL/4DHvgTjEdMvHjwzexCV5nCY5tuZ95+P9958zqdjuk7dhgxtj80fG7k/IWLo5e0y1euXhsbn7i+EnrtwOIVy3O8oGoaIXdsl1ciO3J41Q+40TIdvmpuPRXPVzs8CG3PXY52fb7RMhquXbctI8JQKdocz7JcocAeFO5njjfyOSavLCTXojcx8gV0qIEHFrShBRxciLDtgAEhvtYhDwx8jG1AD2MBtmz5nMMeaMi2UcVRYWB0C+8N7K0nURf7wjOUtIWjOPgOkMzANPvOPrED9o19Zj/Yn4FePekhctnFTzNmub859vJG+fepVAs/I2j2qRNzjqAOj2SuNubuy4iYhRXzne7bg/LjpenebfaB/cT837N99hVn4HZ+WR9LfOkdugt/F6nncr4tmYGLFe5hXNSvISM76Cgih/l5OJroBxjJJLoXR0od62Zjz0ZtmNT9tDHEDGhjxMrjY2hyZXCsiQ4dzCP2MtEnbpvy+6wd+WdgCp9MIbv3H9smsW0lG5DYQMmaJNZUsj6J9ZXsNondVrIhiQ2VbJPENpXsPImdV7KzJHZWyc6R2Dklu0BiF5TsGoldU7JVEltVsjMkdkbJdrEVoJriwJQOntyPG6ijeOgpvXqt2sl+RnPr61VuQsfxTnXr61VujvwHMDFC9UsTg6pPz657Qm47WAeai1CqHM6SyeA8WsSdWZdKlUND1oy+Avp69ax8uUbEKcElzy7NqFyXU+sq9qkjq6fix5kaafSacrxy6leRHq8fjxkNT2uHR7LM4MbK3Vye5fKle9nik+TcNgo34RbcwbPZQyjinrwIFTwJcHgFr+GNNqkVtKKWaIeHEmYS/rm0Z38BNcAYCQ==</latexit>
<latexit

<latexit sha1_base64="ryuXTQ0mJCzzWoku/BcVPGMW6NI=">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</latexit>

T obs = {0, t1 , t2 , . . . , tn = T }

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

13/40

13.

4次元変分法 Adjoint モデル d dt t = (r x f )> X t + (t t0 2T obs <latexit sha1_base64="1kqcoTy2iTY3iSE4F+wj0lrXbFY=">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</latexit> t0 )r xt0 C <latexit sha1_base64="h5FlaEbyb5cnhAtKyCqna4vs1GU=">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</latexit> T =0 4次元変分法による初期値最適化 C(✓) <latexit sha1_base64="+Bo7cbQ29uvKFaPe0edTW1U+aGQ=">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</latexit> ⓪ 初期推定値を設定 ① シミュレーションモデルを解く ② Adjoint モデルを時間後ろ向きに解く ✓guess <latexit sha1_base64="YnLegehdSrKrSmzbu3/KDfP7jUU=">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</latexit> ✓ˆ <latexit sha1_base64="e8y7Sn9k5Dcdy/FYFy9xaHHxaEg=">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</latexit> ✓ <latexit sha1_base64="jRRHIl4aQaDCLy6WjaPD1Ye14kE=">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</latexit> ③ 勾配に従って少し動かす。①へ戻る。 Adjointモデルの計算コスト ≒ シミュレーションモデルの計算コスト 全体の計算コスト = (シミュレーション時間) x (反復回数) c.f. 数値微分 (シミュレーション時間) x (変数の数) x (反復回数) 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 14/40

14.
[beta]
4次元変分法の適用:反応拡散系
(x,
t)
の時間発展

Allen-Cahn モデル (phase- eld モデル)

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

2つの相の時空間発展を記述

+ (1

<latexit sha1_base64="ez+mHNjUXJcK1+H9U215fb4s8aM=">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</latexit>

<latexit sha1_base64="3qMrFlKRdHaO/yZUGFY7VBmQIxc=">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</latexit>

@m
=0
@t

(x, t)
<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

1

(x, t) : Phase-

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

)

1
+m
2

eld 変数

◆
相2

相1

200格子点

@
⌧
= ✏2 4
@t

✓

相1

(x, t)

0

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

m : 相境界移動方向を決めるパラメータ

300格子点

<latexit sha1_base64="cvTWDqnGB0WWzH7k4gKGdgAcjgM=">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</latexit>

TDGL方程式、結晶成長モデル、延焼モデル、…

x=✏
時間微分はオイラー法で離散化、
タイムステップ: t = 0.1⌧
m = 0.1
ラプラシアンは中心差分で離散化、格子間隔:
<latexit sha1_base64="HLR6k1x6qbu/qB5Y1UYnolx6upE=">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</latexit>

<latexit sha1_base64="YJJkoD+q5W19iUYhRhP1ccTq37A=">AAAJHniclZZda9NQGMefbb7E+rJu3gjeBLuJV+VkCoogDC1ju9vWdStbRknS0zYsTUJeSteyL+AX8EIQFUTFjyGIN+KV4D6CeDlBEC98chpsXE+3x5TmPHnO8/uf/zkJyTF9xw4jxg4nJqfOnD13XrmQu3jp8pXp/MzsZujFgcUrlud4QdU0Qu7YLq9EduTwqh9wo206fMvce5T0b3V4ENqeuxHt+3y3bTRdu2FbRoSpWn5WL3EnMtRIfaCyoqZHRlzLF1iRiUMdDbQ0KEB6rHozU79Ahzp4YEEMbeDgQoSxAwaE+NsBDRj4mNuFPuYCjGzRz+EAcsjGWMWxwsDsHp6beLWTZl28TjRDQVs4ioP/AEkV5tkX9pYdsY/sHfvGfo/V6guNxMs+tuaA5X5t+vG18s9TqTa2EbSG1ImeI2jAPeHVRu++yCSzsAZ8p/fkqHx/fb5/k71k39H/C3bI3uMM3M4P69UaX38q1OvINLDtQBeVdOHaS2NTrEH9ry8V5rBnDvUPjrExiY2lbEBiAylrklhTyvok1peyIYkNpWyLxLak7DKJXZayJRJbkrJLJHZJyq6Q2BUpu01it6VslcRWpWwPowCrKQpMquCJd0kT6ygaeqZe/rzZ2Ncmqw3rZWpJHcczVW1YL1NzxNvLxAxVL0uMW326u94J3rq4DjSVpFKm8D9OxvtoihnT79+wXu7JF3c4+T65ZG9ZRqa6kXkqBjoNZPVMfpSpk0avS8crZ57p7HjD/IDJ4T5BO74rGA02F4ra7eLC2p3C4sN0x6DAdbgBt3BXcBcW8a24ChX8OnbhGbyGN8pz5YPySfk8KJ2cSJmr8M+hfP0DyN7UGQ==</latexit>

問題設定

<latexit sha1_base64="RloqgP/JWQSPgqFK7OeGGGkNOFQ=">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</latexit>

Noisy なスナップショットの時系列データから

(x, 0)
の初期値
(x,
t)
・ m の値
・

<latexit sha1_base64="2wtzSLXCd4Y9W9XjDFPJSpTb80M=">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</latexit>

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

<latexit sha1_base64="cvTWDqnGB0WWzH7k4gKGdgAcjgM=">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</latexit>

を4次元変分法によって推定(60,001次元)
詳細な設定
最適化法: L-BFGS-B 法

fi

fi

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

15/40

15.

4次元変分法の適用:反応拡散系 (x, 0) m の推定 の推定 <latexit sha1_base64="cvTWDqnGB0WWzH7k4gKGdgAcjgM=">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</latexit> <latexit sha1_base64="2wtzSLXCd4Y9W9XjDFPJSpTb80M=">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</latexit> 真値 Iteration 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 16/40

16.

4次元変分法のまとめ 良いところ ・モデルの初期値に関する事後分布を評価 ・事後確率最大解(MAP解)のみを探索 → 大規模モデルのパラメータ推定・初期値問題に非常に強い 改善したいところ ・多峰性による局所解への落ち込み ・通常の4次元変分法では抜け出せない ・より高速な探索 ・モデル計算回数はなるべく少なくしたい C(✓) <latexit sha1_base64="+Bo7cbQ29uvKFaPe0edTW1U+aGQ=">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</latexit> ✓ˆ <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> サンプリングの手法と組み合わせてみる 2021.01.22 スマートサンプリング講演会 ✓guess ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">AAALiHicbdbLctowFAZgp9c0lDZpl90wJYtkEQbSTpPuEnK/kwuXJGIyspFBwbIdW4DB4z5Dt+2b9W1qDJNjfOwN0vn0y+DBklTb4K4sFv/NvXj56vWbt/PvFjLvsx8+Li59qrlWz9FYVbMMy2mo1GUGN1lVcmmwhu0wKlSD1dXuztjrfea43DJv5NBmTUHbJte5RmVYqhHZYZI+LOaLhWJ05XCjNG3klelVeVhaMEjL0nqCmVIzqOvel4q2bPrUkVwzWJAlPZfZVOvSNrsPmyYVzG360dcNcnH1qXDdoVCDZFFQ2UnMI/XNps9NuyeZqc0GpKdbpnSDbJaYbKBZQlCz5RPqtAX1Ap+MZ7NsnzgiF9Z+jYvE4IKPIyjBzZREWHxOZEmL6aTPtJV8aTUcqlqeT1TLaI1DueV8aTl4HkSx0zFPVMWqgmpYNdAW1hYow8pAdaw6aBtrG7SDtQPKsXLQR6yPoF2sXVADqwEqsApQE6sJamG1QG2sNugT1idQB6sD6mJ1QSVWCdrD2gPtY+2DDrAOQD2sHugQ6xB0hHUUexm2MW9DuIy1DLqDdQd0F+su6B7WPdB9rPugB1gPQA+xHoIeYT0CPcZ6DHqC9QT0FOsp6BnWM9BzrOegF1gvQCtYK6CXWC9Br7BegV5jvQa9wXoDWsVaBa1hrYHWsdZBG1gboLdYb0HvsN7FXoYRc1IWnmJsWRKsnbJ9kKgeW0V4uMmlDIvqseVkvNOnDIvqsVU2PDu00sZNIPaqp083mpnNS9kOiBfbENInmZlDpCx1RMQWuzYVqQ8gqsduZbvcsFJ2AjKV2L8telbhGD1Hojb8S6OHPZGoHSTPHI6wTBaMP1dIxw0PKMwvlDaZCPhsdzVI5uTASuSKUY7P9lOCHYdNb/k8dm0aTvZxWg/Pkunhfqz/I/XOOu8nf2s0sj/bxUGXe7O5tUQw9n2z4fm0lDyN4kZtvVD6Vli//J7fKk9PqvPKF+WrsqKUlA1lSzlUKkpV0ZRH5bfyR/mbWcgUMxuZn5OhL+ammc/KzJUp/we42Ss6</latexit> <latexit sha1_base64="YnLegehdSrKrSmzbu3/KDfP7jUU=">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</latexit> © 伊藤伸一 19/40

17.

Langevin dynamics C(✓) <latexit sha1_base64="+Bo7cbQ29uvKFaPe0edTW1U+aGQ=">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</latexit> ・通常の4次元変分法 ✓k+1 = ✓k <latexit sha1_base64="f6KOAzZEms1U/q1sOhWGBj/01GI=">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</latexit> ⌧ r ✓k C + p 2⌧ ⇠k ✓ˆ ✓guess <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="YnLegehdSrKrSmzbu3/KDfP7jUU=">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</latexit> ・Metroplis-adjusted Langevin algorithm ✓k+1 = ✓k <latexit sha1_base64="f6KOAzZEms1U/q1sOhWGBj/01GI=">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</latexit> ⇠k  1 0 q(✓ | ✓) / exp k✓ 4⌧ 0 <latexit sha1_base64="eEBkExoNFEFGoKyPmWBxb01q+EA=">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</latexit> =⇥ k <latexit sha1_base64="+Bo7cbQ29uvKFaPe0edTW1U+aGQ=">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</latexit> p @J k ⌧ + 2⌧ ⇠ @⇥k 2⌧ ⇠k :ガウスノイズ <latexit sha1_base64="u6nlPr36NJQ/+xmwTlGbw6QEYcY=">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</latexit> 提案分布 ⌧ r ✓k C + p C(✓)⇥ k+1 ✓ + ⌧ r✓ Ck22 ⇥ˆopt ✓ ⇥ini <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">AAALjnicbdbLctowFAZgJ72lobRJu+yGKVkkizCQdtJsOs095E4uEJKIychGBgXLdmwBBo/7GN22r9W3qTFMjvGxN0jn0y+DB0tSbYO7slj8NzP74uWr12/m3s5n3mXff1hY/Fhzra6jsapmGZZTV6nLDG6yquTSYHXbYVSoBrtROzsjv+kxx+WWeS0HNmsI2jK5zjUqwxIhbSp9IttM0uBhIV8sFKMrhxulSSOvTK7Kw+K8QZqW1hXMlJpBXfe+VLRlw6eO5JrBgizpusymWoe22H3YNKlgbsOPvnSQi6tPhesOhBoki4LKdmIeqW80fG7aXclMbTogPd0ypRtks8Rkfc0SgppNn1CnJagX+GQ0m2X7xBG5sPZrVCQGF3wUQQlupiTC4nMiS5pMJz2mLedLK+FQ1fJ8olpGcxTKLeVLS8HzIIqdjnisKlYVVMOqgTaxNkEZVgaqY9VBW1hboG2sbVCOlYM+Yn0E7WDtgBpYDVCBVYCaWE1QC6sFamO1QZ+wPoE6WB1QF6sLKrFK0C7WLmgPaw+0j7UP6mH1QAdYB6BDrMPYy7CFeQvC21i3QXew7oDuYt0F3cO6B7qPdR/0AOsBaBlrGfQQ6yHoEdYj0GOsx6AnWE9AT7Gegp5hPQM9x3oOWsFaAb3AegF6ifUS9ArrFeg11mvQKtYqaA1rDfQG6w1oHWsd9BbrLegd1rvYyzBkTsrCU4wtS4K1UrYPEtVjqwgPN7mUYVE9tpyMtvqUYVE9tsqGJ4hm2rgxxF719OmGU7N5KdsB8WIbQvokU3OIlKWOiNhi16Ii9QFE9ditbJcbVspOQCYS+7dFzyoco+dI1IZ/afSwxxK1g+SZwxGWyYLR5zJpu+EBhfmF0gYTAZ/urgTJnOxbiVwxyvHpfkqw7bDJLZ/Hrk7CyT5O6+GJMj3ci/XXU++s817yt0Yje9NdHHS5N51bTQRj3zcbnk9LydMobtTWCqWvhbWLb/nN7clJdU75rHxRlpWS8l3ZVMpKRakqmmIrv5U/yt/MQmY98yPzczx0dmaS+aRMXZnyf3FaLgc=</latexit> ✓guess ⇥ ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="YnLegehdSrKrSmzbu3/KDfP7jUU=">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</latexit> ・4次元変分法 → 局所解からの脱出 ・サンプリング法 → 勾配によるブースト 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 20/40

18.
[beta]
例題
・特徴的な粒サイズ

PN

2
S
i=1 i (t)
K(t) = PN
i=1 Si (t)
Z
where Si (t) =
dx i (x, t)
<latexit sha1_base64="1wEV9IUNAwT8cNYUT6D8u14HT0c=">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</latexit>

<latexit sha1_base64="8MrlNqZIWqkGrmsPjsYwfxPkPCc=">AAAJMHiclZZNaxNBGMeftr6s8aWpXgQvwaTSgoRJFRRFKFpKe0ubpg3tlrAvk2ToZnfZnYQ0IV+gX8CDJwUP4kW86dWLX0CwJ/GoHisI4sFnJ4tZm0kdN2Tn2Wee33/+M7vsjuk7LOSEHE5MTp06feasdi51/sLFS9PpmcubodcKLFq2PMcLKqYRUoe5tMwZd2jFD6jRNB26Ze49ivq32jQImedu8H2f7jaNustqzDI4pqrpXKnaY/05Pv9AZy7P2Hq7o9/P6MWQiTxe3uTz1XSW5Ik4MqNBIQ6yEB9Fb2bqJ+hggwcWtKAJFFzgGDtgQIi/HSgAAR9zu9DDXIARE/0U+pBCtoVVFCsMzO7huY5XO3HWxetIMxS0haM4+A+QzMAs+UBekCPynrwkX8ivsVo9oRF52cfWHLDUr04fXC39+CfVxJZDY0id6JlDDe4Krwy9+yITzcIa8O3u46PSvfXZ3g3yjHxD/0/JIXmHM3Db363na3T9iVC3kalh24YOKunCtRfHplgD+4+vDOSwJ4f6/WNsS4ltSdlAiQ2krKnEmlLWV2J9KRsqsaGUbSixDSm7osSuSNklJXZJyi4rsctSdlWJXZWy20rstpStKLEVKdvFKMBqFQUiVfDEu6SOdSoaeqJe/rwx7Gsqqw3rZWpRHcWzqtqwXqbmiLeXiRlVvSQxbvXV3XVP8NbBdVBTiSplCv/jZLyPupix+v0b1ss9+eIOR98nV9lbkpGpbiSeioFODVk9kR9lbKXRbel4pcQznRxvmB8wKdwnFI7vCkaDzYV84VZ+Ye12dvFhvGPQ4BpchzncFdyBRXwrFqGMX8cDeAVv4K32WvuofdI+D0onJ2LmCvx1aF9/AxGt3AU=</latexit>

・擬似観測データ

K(t) = K(t) + !t
K(t) = K(t) + !t :ガウスノイズ
<latexit sha1_base64="WW5GC4d0bdXObbU07f3BItN9xp8=">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</latexit>

<latexit sha1_base64="WW5GC4d0bdXObbU07f3BItN9xp8=">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</latexit>

106

short time window
long time window

5

10
Cost function J

コスト関数 104

・コスト関数がガタガタになることがある
→ 通常の4次元変分法が使えない

103
102
101
100

✓k+1 = ✓k

-1

10

<latexit sha1_base64="f6KOAzZEms1U/q1sOhWGBj/01GI=">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</latexit>

0

0.5

1
1.5
A true
A/A

2

2021.01.22 スマートサンプリング講演会

⌧ r ✓k C +

p

2⌧ ⇠k

ノイズの力を借りて局所解から脱出

© 伊藤伸一

22/40

19.

レプリカ交換モンテカルロ法 逆温度 温度の異なる分布族 i C(✓) <latexit sha1_base64="+Lovzm6f5E7MX0IcjbHLpQTmVZQ=">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</latexit> 高温 山越えが容易 <latexit sha1_base64="NZMBtG1Ob0ZGYghUjAHnEqZZcdw=">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</latexit> p i (✓ | D) / e 0= <latexit sha1_base64="XaYSvgD54zRdFUdMkBhR6M0aml0=">AAAJM3iclZbPaxNBFMdfW3+s8UdTvQhegmnVU5iNgiIWipbS3tqmaUObEvbHJFm6v9jdhDYh/4Dg2YMnBQ/iUbzpyYv/gIfiTbyIeKogiAffTlazNpP2uSE7b9+8z3e+M7vsju7bVhgxtj82PnHi5KnTypnM2XPnL0xmpy6uh14rMHjZ8GwvqOhayG3L5eXIimxe8QOuObrNN/SdB3H/RpsHoeW5a9Gez7cdreFadcvQIkzVstfZbFXnkVbrqr17SVTEyPSi8M+125tVa9k8KzBx5IYDNQnykBzL3tTET6iCCR4Y0AIHOLgQYWyDBiH+tkAFBj7mtqGLuQAjS/Rz6EEG2RZWcazQMLuD5wZebSVZF69jzVDQBo5i4z9AMgcz7AN7wQ7Ye/aSfWG/Rmp1hUbsZQ9bvc9yvzb58HLpx7GUg20EzQF1pOcI6nBHeLXQuy8y8SyMPt/uPD4o3V2d6V5jz9hX9P+U7bN3OAO3/d14vsJXnwh1E5k6tm3YRaWqcO0lsS7WwPzrKwfT2DON+r1DbIvEtqRsQGIDKauTWF3K+iTWl7IhiQ2lbJPENqXsIoldlLLzJHZeyi6Q2AUpu0Ril6TsJondlLIVEluRsh2MAqymKDCpgifeJQ2so2hUU/Xy583CPoesNqiXqcV1HM9UtUG9TM0Wby8dM1S9NDFq9enuOkd428V1oKnElTKF/3Ey2kdDzJh+/wb1ck++uMPx98kle0szMtW11FPR16kjW03lhxmTNLopHa+UeqbT4w3yfSaD+wT18K5gOFgvFtSbheLKrfzc/WTHoMAVuAo3cFdwG+bwrbgMZfw6PoJX8AbeKq+Vj8on5XO/dHwsYS7BP4fy7TfKjt3C</latexit> 0= <latexit sha1_base64="XaYSvgD54zRdFUdMkBhR6M0aml0=">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</latexit> 1 < 1 < 2 2 < ··· < < ··· < n n =1 = 1 が元々の事後分布 1. それぞれは独自にマルコフ連鎖で更新 2. 時々異なる温度間でサンプルを交換 単純なメトロポリス法では超えられない 大きな障壁を越えることが可能 山越えが困難 低温 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 23/40

20.

レプリカ交換モンテカルロ法 逆温度 温度の異なる分布族 i C(✓) <latexit sha1_base64="+Lovzm6f5E7MX0IcjbHLpQTmVZQ=">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</latexit> 高温 山越えが容易 <latexit sha1_base64="NZMBtG1Ob0ZGYghUjAHnEqZZcdw=">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</latexit> p i (✓ | D) / e 0= <latexit sha1_base64="XaYSvgD54zRdFUdMkBhR6M0aml0=">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</latexit> 0= <latexit sha1_base64="XaYSvgD54zRdFUdMkBhR6M0aml0=">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</latexit> 1 < 1 < 2 2 < ··· < < ··· < n n =1 = 1 が元々の事後分布 1. それぞれは独自にマルコフ連鎖で更新 2. 時々異なる温度間でサンプルを交換 単純なメトロポリス法では超えられない 大きな障壁を越えることが可能 山越えが困難 低温 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 24/40

21.

数値実験 Metropolis-adjusted Langevin algorithm + レプリカ交換モンテカルロ法 異なる初期値 B/Btrue Log(コスト関数) A/Atrue 初期値に依らず最適解近くに高速に到達 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 25/40

22.

前半のまとめ 4次元変分法 ・モデルの初期値に関する事後分布のMAP解を高速に評価 ※良い initial guess を選択できれば 大規模モデルのパラメータ推定・初期値問題に非常に強い ・サンプリング法との融合可能性 ・Metroplis-adjusted Langevin algorithm ・Hamiltonian/Hybrid Monte Carlo (HMC) ← レプリカ交換法との併用 ・Hamiltonian/Hybrid Monte Carlo (HMC) v2 H(x, v) = − log(P(x)) 2 dv ∂H =− dt ∂x e v H(x, v) dx ∂H = dt ∂v H(v,x) からのサンプリング <latexit sha1_base64="BhXbVZTemVHN/T08lXChJUPdUqA=">AAALjnicbdZbU5tAFAdwtDdrmlbbx75kGh90pjrBdqwvnXo33q8xXtY6C1mSVRYQNohh6Mfoa/u1+m1KSMYDHHjJ7v7Of0mYsLuaY3JP1mr/RkafPX/x8tXY6/HSm/LbdxOT7888u+vqrKHbpu2ea9RjJrdYQ3JpsnPHZVRoJmtqd6t9b/rM9bhtncpHh10L2ra4wXUq4yHCfoaz9Wn/czATjd9MVGtzteSq4IY6bFSV4XV4MzlukpatdwWzpG5Sz7tSa468DqkruW6yqEy6HnOofkfb7CpuWlQw7zpMvnRUSWtIhec9Ci3KDwoqO7l5pLF4HXLL6Upm6dmADAzbkl5ULhOLPei2ENRqhYS6bUGDKCT92WwnJK6oxGO/+oPE5IL3IyjBrYJEPPiUKJMWM4jP9OmqOhOXanYQEs02W/1QZaqqTkVPRRQ77fNANawaqI5VB21hbYEyrAzUwGqAtrG2QTtYO6AcKwe9xXoLeof1DtTEaoIKrALUwmqB2lhtUAerA3qP9R7UxeqCelg9UIlVgnaxdkF9rD7oA9YH0ABrAPqI9RG0h7WXehmWMS9DeAXrCugq1lXQNaxroOtY10E3sG6AbmLdBK1jrYNuYd0C3ca6DbqDdQd0F+su6B7WPdB9rPugB1gPQA+xHoIeYT0CPcZ6DHqC9QT0FOspaANrA/QM6xloE2sT9BzrOegF1gvQS6yXqZehx9yChaeWWpYEaxdsHyQZT60iPN7kCsqS8dRy0mGyqCwZT62y8QmiVVQ3gNSrXjxdLzNbULAdkCC1IRRPkplDFCx1RKQWuzYVhQ8gGU/dyvG4aRfsBGQoqX9b8qziGqNCkjb8S5OHPZCkHeXPHK6wLRb1P6dJx4sPKCycUxeZiHi2OxPlc/LBzuVqSY5n+wXBjsuGt3yqnR2G832cNuITZXHYT/UXCu9scD//W5NKP9vFQY8H2dxsLpj6vuX4fKrmT6O4cTY/p36Zmz/6Wl1aGZ5Ux5SPyidlWlGVb8qSUlcOlYaiK47yW/mj/C1NlBZK30s/BqWjI8PMByVzler/ATcELKk=</latexit> → 採択率が向上 2021.01.22 スマートサンプリング講演会 最適化法とサンプリング法 両者の相補的な高度化 x © 伊藤伸一 26/40

23.

後半の話題 もっと改善したいところ ・4次元変分法単体では事後分布の形状評価は原理的に不可能 ・分散くらいは評価できないか? → 推定の不確実性・信頼性評価 <latexit sha1_base64="aKB9kaQCOyKzH6yU8shDDgcB/1o=">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</latexit> ・サンプリング法との組み合わせ? ・ 計算回数をなるべく減らしたい ラプラス近似による不確実性評価 Ito et al. (2016) p(✓ | D) ✓ˆ ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 27/40

24.

大規模不確実性評価 © 伊藤伸一 28/40

25.

後半の話題 もっと改善したいところ ・4次元変分法単体では事後分布の形状評価は原理的に不可能 ・分散くらいは評価できないか? → 推定の不確実性・信頼性評価 <latexit sha1_base64="aKB9kaQCOyKzH6yU8shDDgcB/1o=">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</latexit> ・サンプリング法との組み合わせ? ・ 計算回数をなるべく減らしたい ラプラス近似による不確実性評価 Ito et al. (2016) p(✓ | D) ✓ˆ ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 29/40

26.

ラプラス近似 1 ˆ + (✓ C(✓) ⇠ C(✓) 2 ˆ > H✓ˆ (✓ ✓) ˆ ✓) > H✓ˆ = r✓ˆ r✓ˆ C <latexit sha1_base64="JRv58tOCfHzabxlnMrLO7yysm+M=">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</latexit> <latexit sha1_base64="aKB9kaQCOyKzH6yU8shDDgcB/1o=">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</latexit> p(✓ | D) ⇠ <latexit sha1_base64="A7asVjfc253gxyZu5hGmgX7o5lw=">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</latexit> s det H✓ˆ exp N (2⇡) " 1 (✓ 2 : ヘッセ行列 <latexit sha1_base64="GH4BPXo9EmMbY4bo5jJTTNEfX4E=">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</latexit> ˆ✓)> H✓ˆ (✓ ˆ ✓) # p(✓ | D) H✓ˆ 1 <latexit sha1_base64="0lZJ2vqCh+UO6B0zkPKv0KowVfM=">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</latexit> ✓ˆ ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> → ヘッセ行列/逆行列の大きさは (変数の数)2 → 直接法に基づいてヘッセ逆行列を計算しようとすると 計算コスト = (シミュレーション時間) x (変数の数)2 + (変数の数)3 要素計算 逆行列計算 ヘッセ行列/逆行列の直接計算は避けたい 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 30/40

27.
[beta]
クリロフ部分空間法
ヘッセ逆行列要素を部分的に抽出することを考える

Hy = b

b = ei : 単位行列の列ベクトル に取れば
<latexit sha1_base64="dsIv2iRo2y2HFt6e3ayLnTAlzZ0=">AAALiXicbdZbb9owFAfwdNeujK3dHveCRh/ahyLSTVs1aVJber/SC5e2QZUTHHCJkzQxEIiy77DX7ZPt2ywE1JPkJC/Y/p2/AxGxrdoGc0W5/G/u2fMXL1+9nn+zkHubf/d+celD3bX6jkZrmmVYTlMlLjWYSWuCCYM2bYcSrhq0ofYqE28MqOMyy7wWI5u2OOmYTGcaEeFQQ/1J730W3C8Wy6VydBVwQ541itLsqt4vLRhK29L6nJpCM4jr3sllW7R84gimGTTIK32X2kTrkQ69C5sm4dRt+dH3DQpx9Ql33RFXg/QgJ6KbmkfoGy2fmXZfUFNLBoSnW6Zwg3xeMelQszgnZttXiNPhxAt8ZTKbZfuKwwvh2K/JoGIwziYRlGBmRiIcfErklTbVlQHVVoryaliqWp6vqJbRnoQKy0V5OXgqItjJhKeqYlVBNawaaBtrG5RipaA6Vh20g7UD2sXaBWVYGegD1gfQHtYeqIHVAOVYOaiJ1QS1sFqgNlYb9BHrI6iD1QF1sbqgAqsA7WPtgw6wDkCHWIegHlYPdIR1BDrGOo69DFuYtyC8jXUbtIK1ArqDdQd0F+su6B7WPdB9rPugB1gPQA+xHoIeYT0CPcZ6DHqC9QT0FOsp6BnWM9BzrOegVaxV0AusF6CXWC9Br7BegV5jvQatYa2B1rHWQRtYG6BNrE3QG6w3oLdYb2Mvw5g6GQtPObYscdrJ2D6UaDy2irBwk8soi8Zjy0mXiqyyaDy2yoaHh3ZW3RRir3r2dOPEbF7GdqB4sQ0he5LEHDxjqVN4bLHrEJ75AKLx2K1slxlWxk6gzCT2b4ueVVijF5SoDf/S6GFPJWoH6TOHwy2TBpPPFaXrhgcU6pfkDcoDluyuBumcGFqpXDnKsWQ/I9h16OyWT7Vrs3C6j9N6eJjMDg9i/W+Zd9bZIP1bo8pBsouDLvOSubVUMPZ98+H5VE6fRnGjvl6Sv5TWL74WN7dnJ9V56ZP0WVqRZOm7tCkdSFWpJmlST/ot/ZH+5nI5ObeR+zEtfTY3y3yUEleu8h95ByuW</latexit>

<latexit sha1_base64="Rm0o21olSgYXQrfkXukYcHF+pms=">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</latexit>

<latexit sha1_base64="XXSzyGhRQ03wISrqGMzgv7kc/cM=">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</latexit>

クリロフ部分空間法

⇣
⌘
yi = H✓ 1

⇣
⌘
1
y j = H✓
は共分散
は分散、
i, j
i,i

2

<latexit sha1_base64="U97LctVi2kH3/765A/fd+0Zg4sw=">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</latexit>

<latexit sha1_base64="6VrghFAwfQEineCDUlE0dA92+Mo=">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</latexit>

・共役勾配(CG)法

r 1

Kr (H, b) = span{b, Hb, H b, . . . , H

b}

・共役残差(CR)法
・一般化最小残差(GMRES)法,...

CR法の疑似コード
r0
Hr 0 , p0 = r 0
Compute s0 = b
Set an initial guess
For k = 0, 1, 2, ...

↵k = (Hsk , sk ) / (Hsk , Hsk )
r k+1 = r k + ↵k pk
sk+1 = sk ↵k Hpk
k = (Hsk+1 , sk+1 ) / (Hsk , sk )
pk+1 = sk+1 + k pk
Hpk+1 = Hsk+1 + k Hpk

・ヘッセ行列- ベクトル積が高速に計算できれば
逆行列要素が高速に抽出可能になる

End For

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

31/40

28.

Second-order adjoint (SOA) 法 シミュレーションモデルおよび Adjoint モデルを各変数について線形化する シミュレーションモデル x0 = ✓ <latexit sha1_base64="3tnKdZhOroykKE9AkYQLgeiHt+E=">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</latexit> Adjoint モデル d dt = (r x f ) t > t + X (t t0 2T obs <latexit sha1_base64="1kqcoTy2iTY3iSE4F+wj0lrXbFY=">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</latexit> <latexit sha1_base64="h5FlaEbyb5cnhAtKyCqna4vs1GU=">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</latexit> T 0 t )r xt0 C =0 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 32/40

29.
[beta]
Second-order adjoint (SOA) 法
シミュレーションモデルおよび Adjoint モデルを各変数について線形化する

<latexit sha1_base64="Yn4fesY/G2osg9+bcjgQv8fTPeE=">AAALv3icbdZbb5swFABg2t26Ztna7XEv0dKH9KFR0k1bpWlS7/d7m15dRQYMcYOBgpOQIPZf9mv2uj3u34zQqIdw4CX2+c4xwQLbqmtxX9Zq/yYmnz1/8fLV1Ovpwpvi23czs+8vfKfjaayhOZbjXanUZxa3WUNyabEr12NUqBa7VNtrQ7/sMs/njn0u+y67E9S0ucE1KuNQc+Y7MTyqhXoU6jIiOrMkbYYyKv0oEYsZskJsqlpxKIgM4nGzJechqTlTrlVryVXCjfqoUVZG13FzdtoiuqN1BLOlZlHfv63XXHkXUk9yzWJRkXR85lKtTU12GzdtKph/FyZPGZXSGlLh+32hRtmgoLKVGUcaS3cht92OZLY2XiADw7GlHxWLxGY9zRGC2npIqGcKGkQhGY7muCHxRCmO/RwGicUFH5agCm7nVMTBp4piPHUG6TKtUq7Px6mqE4REdSx9WFSaK9fnoqckip0O+VFVrCqohlUD1bHqoAwrAzWwGqAmVhO0hbUFyrFy0Hus96BtrG1QC6sFKrAKUBurDepgdUBdrC7oA9YHUA+rB+pj9UElVgnawdoB7WLtgvaw9kADrAFoH2sfdIB1kPoYVjCvQPEq1lXQNaxroOtY10E3sG6AbmLdBN3CugW6jXUbdAfrDugu1l3QPax7oPtY90EPsB6AHmI9BD3CegR6jPUY9ATrCegp1lPQM6xnoOdYz0EbWBugF1gvQC+xXoJeYb0CvcZ6DXqD9Sb1MQyYl7Pw1FLLkmBmzvZBknhqFeHxJpeTlsRTy0mLyby0JJ5aZeMjh56X9wipTz1/uMHYaEHOdkCC1IaQP8jYGCJnqSMitdiZVOROQBJP3cr1ueXk7ARkJKm3LZmrOMcokaQNb2ky2Y+StKPsmcMTjs2i4W+FtPz4gMLCan2JiYiPd+ejbJ3sOZm6WlLHx/s5hS2PjW75lLswKs72cbURH0Hzi7up/tfcOxu8m33WJLM73sWFPg/G6xYyhan/W4zPp/XsaRQ3Lhar9c/VxZMv5eXV0Ul1SvmofFIqSl35piwr28qx0lA05ZfyW/mj/C2sFMyCXXAfUycnRjUflLGr0P8PqrVCbw==</latexit>

d
dt

Tangent linear (TL) モデル
t

= (r x f )

t

<latexit sha1_base64="nOOznfNHsRu/VgwTqdVSyfoIVns=">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</latexit>

0

=r

Second-order adjoint (SOA) モデル

d
⇠t = (r x f )> ⇠t + {(r x r x f ) t }>
dt

t +

<latexit sha1_base64="cWgBysKO1kJYknbs8RDoTHXoJrk=">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</latexit>

<latexit sha1_base64="EK7rs6L3wFhb50BjyIGqKd0OB7Q=">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</latexit>

⇠T = 0

<latexit sha1_base64="lFXHXN5wHy1jjc38s9iNFpRQ4pg=">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</latexit>

⇠0 = H✓

0

X

t0 2T obs

(t

t0 )r xt r xt >C

t

= H✓ r

TLモデルとSOAモデルを組み合わせることで、
ヘッセ行列と任意のベクトルの積をベクトルとして出力することができる!
2021.01.22 スマートサンプリング講演会

© 伊藤伸一

33/40

30.
[beta]
0

SOA 法によるヘッセ行列-ベクトル積の計算

<latexit sha1_base64="Yn4fesY/G2osg9+bcjgQv8fTPeE=">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</latexit>

d
dt
<latexit sha1_base64="nOOznfNHsRu/VgwTqdVSyfoIVns=">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</latexit>

TL モデル
t
0

= (r x f )
=r

SOAモデル

d
⇠t = (r x f )> ⇠t + {(r x r x f ) t }>
dt

t
<latexit sha1_base64="cWgBysKO1kJYknbs8RDoTHXoJrk=">AAAMaXicbdZbc9M4FAdwh72xJVla9mUHXjwbZkiHaScBhuVlZ6DlUiiXFpq2gLoZ2ZYTUct2LSV1o9F+z/0A+7RfYm0n7bF97JdI+p2/nLqxJCcOuFT9/j+ta999/8OPP13/eeVGu/PLzdW1W4cymiYuG7pRECXHDpUs4CEbKq4CdhwnjAonYEfO6XbuRzOWSB6FB+oiZieCjkPuc5eqbGi0+u8G8RPqas9oTxmS8pFWxv7TJgHzVY+E1AnoSKfGJwkfT9T6X5qoKL4svF+UEV2vxjnisUDRPLMYIOZypiD7sl4hK/eJnIqsdY/w0D7IC1iqdORIY8xigp7aUPfWr6bPU8bW1f7lzNtwz9Fqt7/ZLy4bNwbLRtdaXnujtZWAeJE7FSxUbkCl/Drox+pE00RxN2CmQ6aSxdQ9pWP2NWuGVDB5oov/h7HLqqmQ8kI4pj4oqJrU5lH+kxPNw3iqWOhWAyr1o1BJ0+mQkJ27kRA09DShyVjQ1GiSzxbFmiTCzsb+zgdJwAXPIyjBw4ZENniV6GSPzicz5va6g/Ws1IlSTZwo8PKQfbc7uGuuiih2mvNCHawOqIvVBfWweqAMKwP1sfqgY6xj0AnWCSjHykG/Yf0Geor1FDTAGoAKrAI0xBqCRlgj0BhrDHqG9Qw0wZqASqwSVGFVoFOsU9AZ1hnoOdZz0BRrCnqB9QJ0jnVeehmeYX4G4S2sW6DbWLdBn2N9DvoC6wvQl1hfgr7C+gp0B+sO6Gusr0HfYH0Duot1F/Qt1reg77C+A32P9T3oB6wfQPew7oHuY90H/Yj1I+gnrJ9AD7AegA6xDkEPsR6CHmE9Aj3Gegz6Getn0C9Yv5RehjlLGhaefmlZEmzcsH2QYry0ivBsk2soK8ZLy8mEqaayYry0yhbnjYa6BZRe9ebp5pXZ0obtIDsclbapxkkqc4iGpY6I0mI3pqLxARTjpVvFkgdRw05AllL6tRXPKqvxbVK04VdaPOyFFG1TP3MkIgqZyT97ZCKzAwrTm4MnTBhe7a6bek6dR7Vcv8jxar8hOEnY8pZXtRvLcL2P0352WG4Oz0r9x4139vms/rcWlbNqFwclT6u5jVqw9H072fl0UD+N4sbhg83Bw80H+4+6T7eWJ9Xr1h3rd6tnDaw/rKfWjrVnDS23tds6a81b+sZ/7bX2b+3bi9JrrWXmV6tytbv/A/27hT0=</latexit>

<latexit sha1_base64="EK7rs6L3wFhb50BjyIGqKd0OB7Q=">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</latexit>

⇠T = 0

<latexit sha1_base64="lFXHXN5wHy1jjc38s9iNFpRQ4pg=">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</latexit>

⇠0 = H✓

0

t +

= H✓ r

X

t0 2T obs

(t

t0 )r xt r xt >C

t

t
<latexit sha1_base64="RWXuDuJHXJvlN8xJDO9R8AqtrUs=">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</latexit>
<latexit

Input

0

=r

Output

= H✓ r

t
<latexit sha1_base64="RWXuDuJHXJvlN8xJDO9R8AqtrUs=">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</latexit>
<latexit

<latexit sha1_base64="ryuXTQ0mJCzzWoku/BcVPGMW6NI=">AAAC3XichVG/SxxBFP5cY2I0iRdtAmnE40IKOd5eAgmCINqk9MedCq4uu+t4Lu7tLDtzh7psaSNiJVikSiCF+B+k1cLCNoV/QrA0kCZF3u4thEQ0s+zMN9+8771v5rlR4CtNdNVj9D7oe/io//HA4JOnz4ZKz4cXlWzHnmh4MpDxsusoEfihaGhfB2I5ioXTcgOx5G7NZOdLHRErX4Z1vROJ1ZbTDP0N33M0U3apWl9LLC22dSJdlaaTVkLj2k7MNJtr6bi1LrXKcJhO1q3ULpWpSvkYvQ3MApRRjFlZuoSFdUh4aKMFgRCacQAHir8VmCBEzK0iYS5m5OfnAikGWNvmKMERDrNbPDd5t1KwIe+znCpXe1wl4D9m5Sgq9I1O6IYu6JS+0687cyV5jszLDq9uVysie2j/xcLP/6pavGps/lHd61ljA+9zrz57j3Imu4XX1Xd2j28WJuYrySv6TNfs/xNd0RnfIOz88L7MifmP9/hx2Uv2YpU7IzS2ub7MO6CQtdL8t3G3wWKtar6p1ubelqemi6b24yXG8Jo79w5T+IBZNLjOEb7iDOeGbewZB8ZhN9ToKTQj+GsYx78BYuOsqg==</latexit>

T obs = {0, t1 , t2 , . . . , tn = T }

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

34/40

31.

SOA 法を用いた不確実性評価法 Hy = b <latexit sha1_base64="Rm0o21olSgYXQrfkXukYcHF+pms=">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</latexit> ⓪ y の初期値を設定 <latexit sha1_base64="aKB9kaQCOyKzH6yU8shDDgcB/1o=">AAACvnichVG7TsMwFD2E97MFFiSWiqoIlsoBJBBTBQyMvNoiUVQlwQWLJI4St2qp+AF+gIEJBAPiM1gYWBn4BMQIEgsDN2kkBAi4luPr43uuj3NMzxaBYuyxTWvv6Ozq7unt6x8YHEokh0cKgaz6Fs9b0pb+tmkE3BYuzyuhbL7t+dxwTJsXzcPl8LxY434gpLulGh7fdYx9V1SEZSiCysmEN1VSB1wZJUfspVamy8k0y7IoUj8TPU7SiGNNJu9Rwh4kLFThgMOFotyGgYDGDnQweITtokmYT5mIzjmO0UfcKlVxqjAIPaTvPu12YtSlfdgziNgW3WLT9ImZQoY9sGv2wu7YDXti77/2akY9Qi0NWs0Wl3vlxMnY5tu/LIdWhYNP1p+aFSpYiLQK0u5FSPgKq8WvHZ2+bC5uZJqT7II9k/5z9shu6QVu7dW6WucbZ3/oMUlL+Mcyv1Yo1Ol+GTkQ4Jis1L8b9zMpzGT12ezM+lw6txSb2oNxTGCKnJtHDqtYQz5y9wyXuNJyWkVzNNkq1dpizii+hFb/AL0EnvQ=</latexit> p(✓ | D) H✓ˆ ① TL モデルを解く 1 <latexit sha1_base64="0lZJ2vqCh+UO6B0zkPKv0KowVfM=">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</latexit> ② SOA モデルを時間後ろ向きに解く ✓ˆ ③ クリロフ部分空間法で更新。①へ戻る。 b = ei に取れば <latexit sha1_base64="dsIv2iRo2y2HFt6e3ayLnTAlzZ0=">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</latexit> <latexit sha1_base64="XXSzyGhRQ03wISrqGMzgv7kc/cM=">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</latexit> ⇣ ⌘ yi = H✓ 1 i,i ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> ⇣ ⌘ 1 は分散、 y j = H✓ <latexit sha1_base64="6VrghFAwfQEineCDUlE0dA92+Mo=">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</latexit> i, j は共分散 TL, SOAモデルの計算コスト ≒ シミュレーションモデルの計算コスト 逆行列の計算コスト = (シミュレーション時間) x (反復回数) x (変数の数) 1列のみの抽出であれば、 (シミュレーション時間) x (反復回数) c.f. 直接法 (シミュレーション時間) x (変数の数)2 + (変数の数)3 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 35/40

32.
[beta]
不確実性評価法の適用
(x,
t)
の時間発展

Allen-Cahn モデル (phase- eld モデル)

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

2つの相の時空間発展を記述

+ (1

<latexit sha1_base64="ez+mHNjUXJcK1+H9U215fb4s8aM=">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</latexit>

<latexit sha1_base64="3qMrFlKRdHaO/yZUGFY7VBmQIxc=">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</latexit>

@m
=0
@t

(x, t)
<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

1

(x, t) : Phase-

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

)

1
+m
2

eld 変数

◆
相2

相1

200格子点

@
⌧
= ✏2 4
@t

✓

相1

(x, t)

0

<latexit sha1_base64="Z6IVwu0a+OQ2hYxc9Gak7CYeHIE=">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</latexit>

m : 相境界移動方向を決めるパラメータ

300格子点

<latexit sha1_base64="cvTWDqnGB0WWzH7k4gKGdgAcjgM=">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</latexit>

TDGL方程式、結晶成長モデル、延焼モデル、…

x=✏
時間微分はオイラー法で離散化、
タイムステップ: t = 0.1⌧
m = 0.1
ラプラシアンは中心差分で離散化、格子間隔:
<latexit sha1_base64="HLR6k1x6qbu/qB5Y1UYnolx6upE=">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</latexit>

<latexit sha1_base64="YJJkoD+q5W19iUYhRhP1ccTq37A=">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</latexit>

問題設定

<latexit sha1_base64="RloqgP/JWQSPgqFK7OeGGGkNOFQ=">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</latexit>

Noisy なスナップショットの時系列データから
・

m の不確実性 m
<latexit sha1_base64="cvTWDqnGB0WWzH7k4gKGdgAcjgM=">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</latexit>

<latexit sha1_base64="FE36sNuOQnl09xcmrmLUqJxOa+c=">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</latexit>

(60,001次元空間中の事後分布の標準偏差1つだけ)

のみをSOA法によって選択的に評価
詳細な設定
事前分布: 一様分布 ノイズ分布:

q(!) = N (0, 0.012 )
<latexit sha1_base64="dI48rqdGMKzo3Qy4xFrpWs/Z14Y=">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</latexit>

データ取得: 時間窓 [0.1⌧, 102.5⌧ ] クリロフ部分空間法:共役残差法
<latexit sha1_base64="hDPnxqQvgfmb4Q46gOzQT/F6bkQ=">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</latexit>

fi

fi

2021.01.22 スマートサンプリング講演会

© 伊藤伸一

36/40

33.

不確実性評価法の適用 データ取得間隔 <latexit sha1_base64="vVmVVIaLWXq2BIlNcVdjI92c70o=">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</latexit> T 102.5⌧ 0.1⌧ <latexit sha1_base64="KZ6ct72X8EumPTb4WWj5AhnZt04=">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</latexit> <latexit sha1_base64="3EirMri1ChkJhn6e12th86FFPoE=">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</latexit> m <latexit sha1_base64="FE36sNuOQnl09xcmrmLUqJxOa+c=">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</latexit> m <latexit sha1_base64="FE36sNuOQnl09xcmrmLUqJxOa+c=">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</latexit> Time は4次元変分法による推定値 / 1/Ndata <latexit sha1_base64="hDUJnfh6dczxkkLGXk1C2Y1Snqo=">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</latexit> 高次元の事後分布からの選択的な不確実性抽出が可能に 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 37/40

34.

まとめと展望 まとめ <latexit sha1_base64="aKB9kaQCOyKzH6yU8shDDgcB/1o=">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</latexit> p(✓ | D) H✓ˆ ・4次元変分法データ同化 1 <latexit sha1_base64="0lZJ2vqCh+UO6B0zkPKv0KowVfM=">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</latexit> ・サンプリング法との相補的な高速化へ向けた試み ・Second-order adjoint 法 に基づく不確実性評価 ✓ˆ ✓ <latexit sha1_base64="6L/sfM20+IbPGS/8f9cvkF/lmnE=">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</latexit> <latexit sha1_base64="ehxCx+d3pIESsKl7ZfcPjJ0tcZ8=">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</latexit> 今後の展望 ・いかに少ない試行で意味のある情報をとってこれるか ・最適化法とサンプリング法の良いとこ取り 2021.01.22 スマートサンプリング講演会 © 伊藤伸一 39/40

35.

ありがとうございました。 © 伊藤伸一 40/40