Transcript
Page 1: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

The Legacy of Rudolph KalmanBlending Data and Mathematical Models

Andrew Stuart

Computing and Mathematical SciencesCalifornia Institute of Technology

AFOSR, DARPA, EPSRC, ONRIsaac Newton Institute, Rothschild Lecture

April 5th 2018

Page 2: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Overview

Historical Context

Kalman State Estimation

Weather Forecasting

Enemble Kalman Inversion

Inversion Applications

Conclusions and References

Page 3: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Historical Context

Page 4: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Newton

Data: conceptual/sparse =⇒ Models: conservation laws.

Page 5: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Einstein

Data: conceptual/sparse =⇒ Models: special and general relativity.

Page 6: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Kalman State Estimation

Page 7: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Kalman Filter

State Space Model

Dynamics Model: vn+1 = Mvn + ξn, n ∈ Z+

Data Model: yn+1 = Hvn+1 + ηn+1, n ∈ Z+

Probabilistic Structure: v0 ∼ N(m0,C0), ξn ∼ N(0,Σ), ηn ∼ N(0, Γ)

Probabilistic Structure: v0 ⊥ {ξn} ⊥ {ηn} independent

I Born: Budapest, May 19, 1930.

I Died: Florida, July 2, 2016.

I BS and MS from MIT, 1953, 1954.

I Positions at Stanford, ETH, U of Florida.

I US National Academy of Engineering 1991.

I US National Academy of Sciences 2004.

I US National Medal of Science 2008.

I Draper Prize, Kyoto Prize, Steele Prize . . .

Page 8: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Kalman Filter

State Space Model

Dynamics Model: vn+1 = Mvn + ξn, n ∈ Z+

Data Model: yn+1 = Hvn+1 + ηn+1, n ∈ Z+

Probabilistic Structure: v0 ∼ N(m0,C0), ξn ∼ N(0,Σ), ηn ∼ N(0, Γ)

Probabilistic Structure: v0 ⊥ {ξn} ⊥ {ηn} independent

I J. Basic Engineering 82(1960); see [1].

I 27,307 Google Scholar citations.

I Navigational and guidance systems.

I Apollo 11.

I Yn = {y`}n`=1.

I vn|Yn ∼ N(mn,Cn).

I (mn,Cn) 7→ (mn+1,Cn+1).

Page 9: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Kalman Filter

Sequential Optimization Perspective

Predict: mn+1 = Mmn, n ∈ Z+

Model/Data Compromise: Jn(m) =1

2|m − mn+1|2Cn+1

+1

2|yn+1 − Hm|2Γ

Optimize: mn+1 = argminm Jn(m).

I | · |A = |A−12 · | for A > 0.

I Updating Cn+1 is expensive: O(d2) storage.

I d the state space dimension.

Page 10: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

3DVAR Filter

State Space Model

Dynamics Model: vn+1 = Ψ(vn) + ξn, n ∈ Z+

Data Model: yn+1 = Hvn+1 + ηn+1, n ∈ Z+

Probabilistic Structure: v0 ∼ N(m0,C0), ξn ∼ N(0,Σ), ηj ∼ N(0, Γ)

Probabilistic Structure: v0 ⊥ {ξn} ⊥ {ηn} independent

I Introduced in UK Met Office.

I Primary proponent: Andrew Lorenc.

I Quart J. Roy. Met. Soc. 112(1986).

I J. Met. Soc. Japan 99(1997).

I {vn} 7→ {vn+1}.

Page 11: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

3DVAR

Sequential Optimization Perspective

Predict: vn+1 = Ψ(vn), n ∈ Z+

Model/Data Compromise: Jn(v) =1

2|v − vn+1|2C +

1

2|y (k)

n+1 − Hv |2Γ

Optimize: vn+1 = argminvJn(v).

I C is a fixed model covariance (not updated sequentially).

I C chosen to have simple, computable, structure (Fourier).

Page 12: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Ensemble Kalman Filter

State Space Model

Dynamics Model: vn+1 = Ψ(vn) + ξn, n ∈ Z+

Data Model: yn+1 = Hvn+1 + ηn+1, n ∈ Z+

Probabilistic Structure: v0 ∼ N(m0,C0), ξn ∼ N(0,Σ), ηj ∼ N(0, Γ)

Probabilistic Structure: v0 ⊥ {ξn} ⊥ {ηn} independent

I Introduced by Geir Evensen.

I J. Geophysical Research 99(1994).

I Motivated by extended Kalman filter; see [2].

I Jazwinski (1970) [3], Ghil et al (1981) [4].

I Original paper in ocean dynamics.

I Used in weather forecasting centres worldwide.

I {v (k)n }Kk=1 7→ {v

(k)n+1}

Kk=1.

Page 13: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Ensemble Kalman Filter

Sequential Optimization Perspective

Predict: v(k)n+1 = Ψ(v (k)

n ) + ξ(k)n , n ∈ Z+

Model/Data Compromise: J(k)n (v) =

1

2|v − v

(k)n+1|

2Cn+1

+1

2|y (k)

n+1 − Hv |2Γ

Optimize: v(k)n+1 = argminvJ

(k)n (v).

I Cn+1 is empirical covariance of the {v (k)n+1}.

I Updating Cn requires only O(Kd) storage.

Page 14: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Weather Forecastingw/Law

Page 15: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Weather Forecasting: Data

Page 16: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Data Fails to Overcome Butterfly EffectKJH Law and AM Stuart, Monthly Weather Review, 2014.

5 10 15 20

10−2

10−1

100

step

3DVAR, ν=0.01, h=0.2

||m(tn)−u+(t

n)||2

tr(Γ)

tr[(I−Bn)Γ(I−B

n)*]

0 1 2 3 4

−0.2

−0.1

0

0.1

0.2

0.3

3DVAR, ν=0.01, h=0.2, Re(u1,2

)

t

m

u+

yn

Page 17: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Theory Backed use of Data Overcomes Butterfly EffectKJH Law and AM Stuart, Monthly Weather Review, 2014.

10 20 30 40 50 60 70

10−1

100

step

[3DVAR], ν=0.01, h=0.2

||m(tn)−u+(t

n)||2

tr(Γ)tr[(I−B

n)Γ(I−B

n)*]

0 2 4 6 8 10 12

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

[3DVAR], ν=0.01, h=0.2, Re(u1,2

)

t

m

u+

yn

Page 18: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

DWD Weather Forecasting: Impact of Mathematics

Ensemble Kalman Filter (red) versus 3DVAR (blue)

forecast lead time in hours

higher skill

lower skill

quality gain

Page 19: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Ensemble Kalman Inversionw/Iglesias and Law

w/Schillings

Page 20: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Inverse Problem

Problem StatementFind u from y where G : U 7→ Y, where U ,Y are Hilbert spaces, η is noise and

y = G(u) + η, η ∼ N(0, Γ).

A.N. Tikhonov (1963) J. Franklin (1970)

Page 21: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Inverse Problem

Dynamical Formulation

Dynamics Model: un+1 = un, n ∈ Z+

Dynamics Model: wn+1 = G(un), n ∈ Z+

Data Model: yn+1 = wn+1 + ηn+1, n ∈ Z+

I yn+1 = y , ηn+1 ∼ N(0, h−1Γ).

I Evensen moved to Statoil.

I Methodology widely used in oil industry.

I Also in groundwater flow.

I Gier Nævdal 2001, 2002.

I Oliver, Reynolds, Liu (2008) [5].

Page 22: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

EKI Algorithm

Reformulate In General State Space Notation

vn = (un,wn),Ψ(u,w) =(u,G(u)

),H(u,w) = w .

Continuous Time Limit [10]

Let NT = h. Then as h→ 0

sup0≤n≤N |u(k)(nh)− u(k)n | → 0

where

u(k) = −K∑`=1

dk,`(u)u(`)

wheredk,`(u) = 〈G(u(k))− y ,G(u(`))− G〉.

Page 23: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Properties of EKI

I Linear Case G(·) = A · .I Least Squares Functional

Φ(u) =1

2‖y − Au‖2

Γ.

I Gradient Structure

du(k)

dt= −C∇Φ(u(k)),

C =1

K

K∑`=1

(u(`) − u

)⊗(u(`) − u

). (1)

Theorem (Gradient Structure) [10]

Algorithm minimizes Φ(·; y) over a finite dimensional subspace defined by thelinear span of the initial ensemble.

Page 24: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Inversion Applicationsw/Chada, Iglesias, Roininen

Page 25: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Groundwater Flow 1

Forward ProblemGiven κ ∈ X := L∞(D;R+) find p ∈ H1

0 (D;R) such that:

−∇ · (κ∇p) = f , x ∈ D,

p = 0, x ∈ ∂D.

Inverse ProblemSet κ = exp(u). Given K linear functionals of the pressure Gk(u) = ok(p),ok ∈ H−1(D;R), find u from noisy measurements y where:

y = G(u) + η, η ∼ N(0, Γ).

Page 26: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Grdounwater Flow 2

0 2 4 60

1

2

3

4

5

6

-1

0

1

2

3

4

5

Figure: True log-permeability.

Parameterization

I Five scalars describe channel geometry.

I One random field describes interior of channel.

I One random field describes exterior of channel.

I Lengthscale and smoothness parameters of both random fields.

Page 27: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Groundwater Flow 3

Figure: Five succesive iterations; ensemble mean

Page 28: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Electrical Impedance Tomography (EIT) 1

Forward ProblemGiven (κ, I ) ∈ L∞(D;R+)× Rm find (ν,V ) ∈ H1(D)× Rm :

−∇ · (κ∇ν) = 0 ∈ D,

ν + z`κ∇ν · n = V` ∈ e`, ` = 1, . . . ,m,

∇ν · n = 0 ∈ ∂D\ ∪m`=1 e`,∫

κ∇ν · n ds = I` ∈ e`, ` = 1, . . . ,m. D

∂D

el

Ohm’s Law: V = R(κ)× I .

Inverse ProblemSet κ = exp(u). Given a set of K noisy measurements of voltage V (k) fromcurrents I (k), and Gk(u) = R

(exp(u)

)× I (k), find u from y where:

y(k) = Gk(u) + η, η ∼ N(0, γ2), k = 1, . . . ,K .

Page 29: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

EIT 2

1

1.5

2

2.5

3

3.5

4

4.5

5

Figure: True Conductivity.

Parameterization

I Continuous level set function.

I Lengthscale of level set function.

I Smoothness of level set function.

Page 30: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

EIT 3

Figure: Five succesive iterations: level set function.

Page 31: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

EIT 4

Figure: Five succesive iterations: thresholded level set function.

Page 32: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Supervised Learning

Inverse Problem

I Data: {(xj , yj)}Nj=1 with xj ∈ X , yj ∈ Y and X ,Y Hilbert spaces.

I Find: G(u|·) : X → Y for parameter u ∈ U consistent with the data.

I Concatenate x, y and G(u|·) :

y = G(u|x) + η

where G(·|x) : U → YN and η is model or data error.

Page 33: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Supervised Learning

Key Issues

I Approximation: design of G(·|xj);

I Optimization: choosing u to fit data {(xj , yj)}Nj=1;

I Stability: ability of G(·|x?) to predict well for out of sample x?.

Architecture, training and generalization.

Page 34: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

MNIST Dataset

LeCun and Cortes 1999.

Page 35: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

MNIST Supervised

Figure: Test Accuracy of Net 1 on MNIST (batched).

J Loss Momentum Randomize y Randomize u

5000 Cross Entropy X X χ

Page 36: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Conclusions and References

Page 37: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Conclusions

I Kalman’s 1960 paper revoltionized applied mathematics.

I Evensen’s 1994 paper introduced a step change in applicability.

I Both state estimation and inverse problems maybe solved.

I Aerospace guidance · · ·I Oceanography, weather forecasting, climate . . .

I Geophysical and medical imaging.

I Machine Learning ?

Page 38: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

References

[1] R. Kalman and R. Bucy.

A new approach to linear filtering and prediction problems

Journal of Basic Engineering, 82(1961), 95–108.

[2] G. Evensen.

Data Assimilation: The Ensemble Kalman Filter.

Springer, 2006.

[3] A.H. Jazwinski.

Stochastic Processes and Filtering Theory.

Academic Press, 1970.

[4] M. Ghil, S.E. Cohn, J. Tavantzis, K. Bube, and E. Isaacson.

Application of estimation theory to numerical weather prediction.

Dynamic Meteorology: Data Assimilation Methods, L. Bengtsson, M. Ghil and E. Kallen, Eds.

Springer, 139–224, 1981.

[5] D.S Oliver, A.C. Reynolds and N Liu

Inverse theory for petroleum reservoir characterization and history matching

CUP 2008

Page 39: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

References

[6] A. M. Stuart

Inverse Problems: a Bayesian perspective

Acta Numerica, 19(May):451-559, 2010.

[7] M. Iglesias, Y. Lu and A.M. Stuart

A Bayesian level set method for geometric inverse problems.

Interfaces and Free Boundaries, 18(2016), 181-217.

[8] N. K. Chada, M. Iglesias, L. Roininen, A. M. Stuart

Geometric and Hierarchical Ensemble Kalman Inversion,

arXiv:1602.1709.01781

[9] M. A. Iglesias, K. Law, A. M. Stuart

Ensemble Kalman method for inverse problems,

Inverse Problems, 29 (4), 045001, 2013.

[10] C. Schillings, A. M. Stuart

Analysis of the ensemble Kalman filter for inverse problems,

arXiv:1602.02020, SIAM Num. Analysis 55(2017).

Page 40: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Whittle-Matern Initial Ensembles

I Create initial ensemble of functions via Gaussian random fields.

I Common choice: Whittle-Matern family

cν,τ (x , x ′) :=21−ν

Γ(ν)

(τ |x − x ′|

)νKν(τ |x − x ′|

).

I Smoothness parameter: ν ∈ R+.I Inverse length-scale parameter: τ ∈ R+.

I Corresponding covariance operator

Cν,τ ∝ τ 2ν(τ 2I −4)−ν−d2 .

I Hierarchical: invert for as ν, τ as well as field itself.

I F. Lindgren, H. Rue and J. Lindstrom, (JRSS-B 73(2011))

Page 41: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Centred vs Non-centred

I Define θ = (α, τ).

I Generate samples u by solving the SPDE

(τ 2I −4)ν+ 1

2d

2 u = τνξ,

where ξ ∼ N(0, I ) is white noise. u = T (ξ, θ).See F. Lindgren, H. Rue and J. Lindstrom, (JRSS-B 73(2011))

I Hierarchical: invert for parameters θ as well as field u.

I Centred approach:

I view (u, θ) as unknowns;I initial ensemble samples P(u|θ)P(θ);I y = G(u) + η.

I Non-centred approach:

I view (ξ, θ) as unknowns;I initial ensemble samples P(ξ)P(θ);I y = G

(T (ξ, θ)

)+ η.

See O. Papaspiliopoulos, G. O. Roberts, and M. Skold, (Statistical Science)22(2007))

Page 42: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Supervised Learning

Φ(u) =1

2‖y − G(u|x)‖2

YN or −N∑j=1

〈yj , log G(u|, xj)〉Y

Algorithms

SGD :u = −∇uΦ(u); u(0) = u0, u∗ = u(T )

EnKF :u(k) = −K∑`=1

dk,`(u)u(`); u(k)(0) = u(k)0 , u∗ =

1

K

K∑`=1

u(`)(T )

Tricks

I Mini-batching.

I Momentum: u(k) + 3tu(k) = −

∑K`=1 dk,`(u)u(`).

I Randomize y and u.

Page 43: The Legacy of Rudolph Kalman - newton.ac.uk€¦ · Inverse theory for petroleum reservoir characterization and history matching CUP 2008. References [6] A. M. Stuart Inverse Problems:

Convolutional Models

Net 1 ∼ 14k Net 2 ∼ 30k

Conv12x3x3MaxPool2x2

Conv12x3x3Conv12x3x3MaxPool2x2

Conv24x3x3MaxPool2x2

Conv24x3x3Conv24x3x3MaxPool2x2

Conv32x3x3MaxPool2x2

Conv32x3x3Conv32x3x3

FC-100 FC-100

FC-10 FC-10

I ReLU: after each block: max(0, x);

I Layer normalization: Ba, Kiros and Hinton 2016. (NIPS)


Top Related