Numerical Methods for PDE-Constrained Optimization

Size: px

Start display at page:

Download "Numerical Methods for PDE-Constrained Optimization"

Branden Wright
6 years ago
Views:

1 Numerical Methods for PDE-Constrained Optimization Richard H. Byrd 1 Frank E. Curtis 2 Jorge Nocedal 2 1 University of Colorado at Boulder 2 Northwestern University Courant Institute of Mathematical Sciences, 2007

2 Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

3 Problem Formulation Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

Problem Formulation Data assimilation in weather forecasting Goal: up-to-date global weather forecast for the next 7 to 10 days 1 If an entire initial state of the atmosphere (temperatures,

4 Problem Formulation Data assimilation in weather forecasting Goal: up-to-date global weather forecast for the next 7 to 10 days 1 If an entire initial state of the atmosphere (temperatures, pressures, wind patterns, humidities) were known at a certain point in time, then an accurate forecast could be obtained by integrating atmospheric model equations forward in time Flow described by Navier-Stokes and further sophistications of atmospheric physics and dynamics (none of which will be discussed here) 1 (Fisher, Nocedal, Trémolet, and Wright, 2007)

5 Problem Formulation In reality: partial information Limited amount of data (satellites, buoys, planes, ground-based sensors) Each observation is subject to error Nonuniformly distributed around the globe (satellite paths, densely-populated areas) (Graphics courtesy Yannick Trémolet)

6 Problem Formulation In reality: partial information Limited amount of data (satellites, buoys, planes, ground-based sensors) Each observation is subject to error Nonuniformly distributed around the globe (satellite paths, densely-populated areas) (Graphics courtesy Yannick Trémolet)

7 Problem Formulation Data assimilation: defining the unknowns Currently in operational use at the European Centre for Medium-Range Weather Forecasts (ECMWF) We want values for an initial state, call it x 0 For a given x 0, we could integrate our atmospheric models forward to forecast the state of the atmosphere at N time points x i = M(x i 1 ), i = 1,..., N (x i : state of the atmosphere at time i) Observe the atmosphere at these N time points (y i : observed state at time i) y 1,..., y N Let x b (background state) be values at initial time point obtained from previous forecast carry over old information

8 Problem Formulation Data assimilation as an optimization problem Define the difference 2 x 0 x b 3 f (x 0 x 1 y 1 ) = x N y N and choose x 0 as the initial state most likely to have given the observed data 1 min f (x 0 ) 2 x 0 2 R = 1 (x 0 x b ) T (R b ) 1 (x 0 x b P )+ 1 N i=0 (x i y i ) T (R i ) 1 (x i y i ) 1 2 f (x 0 ) 2 R 1 : distance measure between observed and expected behavior R = (R b, R 1,..., R N ): background and observation error covariance matrices (choice of these values is a separate, but important issue) In current forecasts, x 0 contains approximately unknowns

9 Solution Techniques Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

10 Solution Techniques Computational issues Data assimilation optimization problem: 1 min f (x 0 ) 2 x 0 2 R = 1 (x 0 x b ) T (R b ) 1 (x 0 x b P )+ 1 N i=0 (x i y i ) T (R i ) 1 (x i y i ) Difficulties include: problem is very large ( x ) objective is nonconvex (nonlinear operators M i ) exact derivative information not available solutions needed in real-time

11 Solution Techniques Current algorithm: nonlinear elimination Given a guess of the initial state x 0 1. Apply operators M i to compute the expected state at the N time points x i = M(x i 1 ), i = 1,..., N and evaluate the objective 1 2 f (x 0 ) 2 R 1 2. Compute a step d toward an improved solution a.... derive sensitivities of x 1,..., x N with respect to x 0 to form J(x 0 ) : Jacobian of f (x 0 ) (Note: can only be done inexactly) b.... solve the quadratic problem min d 1 2 f (x 0 ) + J(x 0 )d 2 R 1

12 Solution Techniques Illustration of solution process

13 Solution Techniques Recent advances: multilevel schemes We are interested in solving the subproblem min d q(d) = 1 2 f (x 0 ) + J(x 0 )d 2 R 1 Conjugate Gradients or Lanczos method is applicable, but not at high resolutions Define a restriction operator S such that ˆx i = Sx i, i = 0,..., N Solve min ˆq(ˆd) ˆd where ˆq is a lower-resolution analog of q Use a prolongation operator S + to map the computed step into the higher-resolution space x 0 x 0 + S +ˆd

14 Solution Techniques Future study: weak constraints Previous formulation assumes that model errors can be neglected However, we can lift this questionable assumption by imposing weak constraints x i M(x i 1 ), i = 1,..., N These approximate equalities can be imposed by creating a penalty term in the objective min 1 2 (x 0 x b ) T (R b ) 1 (x 0 x b ) P N i=0 (x i y i ) T (R i ) 1 (x i y i ) P N i=1 (x i M(x i 1 )) T (Q i ) 1 (x i M(x i 1 )) where Q i are model error covariances Note: state vectors x 1,..., x N become unknowns; problem dimension can increase from to as large as !

15 Solution Techniques Future study: weak constraints The weak constraint formulation min 1 2 (x 0 x b ) T (R b ) 1 (x 0 x b ) P N i=0 (x i y i ) T (R i ) 1 (x i y i ) P N i=1 (x i M(x i 1 )) T (Q i ) 1 (x i M(x i 1 )) can be seen as a step toward constrained optimization 1 min (x 0 x b ) T (R b ) 1 (x 0 x b P ) + 1 N x 0,...,x N 2 2 i=0 (x i y i ) T (R i ) 1 (x i y i ) s.t. x i = M(x i 1 ), i = 1,..., N (Note: in solving the constrained formulation, constraints are satisfied only in the limit)

16 Solution Techniques Challenges in PDE-Constrained Optimization Optimization Method Nonlinear elimination Composite-step methods Full space methods DO versus OD Discretize-Optimize Optimize-Discretize Multilevel Schemes Compute steps at low resolution Optimize at low resolution Inexactness Efficiency Robustness Overview of important optimization questions

17 Solution Techniques Challenges in PDE-Constrained Optimization Optimization Method Nonlinear elimination Composite-step methods Full space methods DO versus OD Discretize-Optimize Optimize-Discretize Multilevel Schemes Compute steps at low resolution Optimize at low resolution Inexactness Efficiency Robustness Methods arising in the data assimilation problem

18 Solution Techniques Challenges in PDE-Constrained Optimization Optimization Method Nonlinear elimination Composite-step methods Full space methods DO versus OD Discretize-Optimize Optimize-Discretize Multilevel Schemes Compute steps at low resolution Optimize at low resolution Inexactness Efficiency Robustness Our contribution to the field...

19 Solution Techniques Computational savings with inexact methods Inexact step computations are... effective for unconstrained optimization 2 min x R n f (x)... effective for systems of nonlinear equations 3 F (x) = 0 or min x R n... necessary for many PDE-constrained problems F (x) 2 (Steihaug, 1983) 3 (Dembo, Eisenstat, Steihaug, 1982)

20 Background and Basics Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

21 Background and Basics Equality Constrained Optimization Minimize an objective subject to mathematical equalities min f (x) x R n s.t. c(x) = 0 (where the constraints c(x) = 0 contain the PDE)

22 Background and Basics Characterizing Optimal Solutions Defining and then the derivatives L(x, λ) = f (x) + λ T c(x) g(x) = f (x) and A(x) = [ c i (x)], we have the first order optimality conditions» g(x) + A(x) T λ = 0 c(x)

23 Background and Basics Method of choice: Sequential Quadratic Programming (SQP) At a given iterate x k, formulate the quadratic subproblem min d R n f k + g T k d d T W k d s.t. c k + A k d = 0 for some symmetric positive definite W k ( xxl 2 k ) to compute a step, or, equivalently, solve the primal-dual system»»» Wk A T k dk gk + A T k λ = k A k 0 Measure progress toward a solution with the merit function δ k c k φ π(x) = f (x) + π c(x), π > 0

24 Background and Basics SQP: for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k A k 0 Set π k π k 1 Line search: endfor δ k φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) c k (It works!)

25 Background and Basics SQP: for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k A k 0 Set π k π k 1 Line search: endfor δ k φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) c k

26 Background and Basics SQP: for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k A k 0 Set π k π k 1 Line search: endfor δ k φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) c k

27 Background and Basics SQP: for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k A k 0 Set π k π k 1 Line search: endfor δ k φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) c k The step is good; Dφ πk (d k ) < 0 Algorithm is well-defined; α k > 0 π will stabilize reducing φ π drives search toward solution of optimization problem

28 Algorithm Overview Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

29 Algorithm Overview Inexactness is necessary An SQP algorithm requires the exact solution of the system»»» Wk A T k dk gk + A T k λ = k A k 0 δ k However, for many PDE constrained problems we cannot form or factor this iteration matrix Alternatively, we can consider the application of an iterative solver which, during each inner iteration, yields»»»» Wk A T k dk gk + A T k λ = k ρk + A k 0 with residuals (ρ k, r k ) δ k Iterative methods only require mechanisms for computing products with W k and A k and its transpose, so the algorithm is matrix-free c k c k r k

30 Algorithm Overview Inexact SQP: Idea: terminate when (ρ k, r k ) is small for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k + A k 0 Set π k π k 1 Line search: endfor δ k c k» ρk φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) r k

31 Algorithm Overview Inexact SQP: Idea: terminate when (ρ k, r k ) is small for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k + A k 0 δ k c k» ρk r k For any level of inexactness: Set π k π k 1 Line search: φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) endfor

32 Algorithm Overview Inexact SQP: Idea: terminate when (ρ k, r k ) is small for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k + A k 0 Set π k π k 1 Line search: endfor δ k c k» ρk φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) r k For any level of inexactness: Step may be an ascent direction for φ π Penalty parameter may tend to

33 Algorithm Overview Inexact SQP: We don t know when to terminate the iterative solver for k = 1, 2,... Compute step:»»» Wk A T k dk gk + A T k λ = k + A k 0 Set π k π k 1 Line search: endfor δ k c k» ρk φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) r k For any level of inexactness: Step may be an ascent direction for φ π Penalty parameter may tend to

34 Algorithm Overview Main idea: inexactness based on models Modern optimization methods work with models For the merit function φ π(x) = f (x) + π c(x), define a model about x k : m π(d) = f k + gk T d + 1 d T W 2 k d + π c k + A k d, For a given d k, we can estimate the reduction in φ π by evaluating mred π(d k ) = m π(0) m π(d k ) = gk T d k 1 d T 2 k W k d k + π( c k r k ) (recall r k = c k + A k d k )

35 Algorithm Overview Termination Tests We may terminate the iteration on»»» Wk A T k dk gk + A T k λ = k + A k 0 if the component d k : δ k c k» ρk r k

36 Algorithm Overview Termination Tests We may terminate the iteration on»»» Wk A T k dk gk + A T k λ = k + A k 0 δ k if the component d k : (1) yields a sufficiently large reduction in the model of the merit function for the most recent penalty parameter π k 1 ; i.e. c k mred πk 1 (d k ) σπ k 1 c k» ρk r k

37 Algorithm Overview Termination Tests We may terminate the iteration on»»» Wk A T k dk gk + A T k λ = k + A k 0 δ k if the component d k : (2) yields a reduction in the linear model of the constraints; i.e., for π k > π k 1 c k mred πk (d k ) σπ k c k» ρk r k

38 Algorithm Overview An Inexact SQP Algorithm (Byrd, Curtis, and Nocedal, 2007) Given parameters 0 < ɛ, τ, σ, η < 1 and 0 < β Initialize x 0, λ 0, and π 1 > 0 for k = 0, 1, 2,..., until convergence Set π k = π k 1 Iteratively solve the system» Wk A T»» k dk gk + A = T k λ» k ρk + A k 0 δ k c k r k Terminate the inner iterations when mred πk (d k ) ρ k σπ k c k max{β c k, ɛ g k + A T k λ k } or r k ρ k Perform line search φ πk (x k + α k d k ) φ πk (x k ) + ηα k Dφ πk (d k ) Set (x k+1, λ k+1 ) (x k, λ k ) + α k (d k, δ k ) endfor ɛ c k β c k π k g T k d k dt k W kd k (1 τ)( c k r k )

39 Algorithm Overview Summary: Our contribution SQP is the algorithm of choice for large-scale constrained optimization Introducing inexactness is a difficult issue and naïve approaches can fail to ensure convergence We present two simple termination tests for the iterative solver, where the level of inexactness is based on models of a merit function The algorithm is globally convergent Numerical experiments are considered next...

40 Optimization Test Problems: Robustness Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

41 Optimization Test Problems: Robustness Question: Is the algorithm robust? 44 problems from standard optimization test sets (CUTEr and COPS) Algorithm A: inexactness based on»» ρk κ gk + A T k λ k, 0 < κ < 1 r k Algorithm B: inexactness based on model reductions (Note: can have κ 1!) c k Algorithm Alg. A Alg. B κ % Solved 45% 80% 86% 100%

42 PDE-Constrained Test Problems: Practicality Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

43 PDE-Constrained Test Problems: Practicality Question: Is the algorithm practical? 2 model inverse problems in PDE-constrained optimization 4 where 1 min Qz (y,z) R n 2 d 2 + γr(y y ref ) s.t. A(y)z = q objective = data fitting + regularization constraints = PDE Example applications: Elliptic PDE groundwater modeling, DC resistivity Parabolic PDE optical tomography, electromagnetic imaging 4 (Haber and Hanson, 2007)

44 PDE-Constrained Test Problems: Practicality Question: Is the algorithm practical? First problem (n = 8192, t = 4096) Elliptic PDE (groundwater modeling, DC resistivity)... many details... Solution obtained in 9 iterations Sample output: Note: κ k = (ρ k, r k ) / (g k + A T k λ k, c k ) k KKT Inner Iter. κ k T. Test new π curr. π new π Overall, average κ k

45 PDE-Constrained Test Problems: Practicality Question: Is the algorithm practical? Second problem (n = 69632, t = 65536) Parabolic PDE (optical tomography, electromagnetic imaging)... many details... Solution obtained in 11 iterations Sample output: Note: κ k = (ρ k, r k ) / (g k + A T k λ k, c k ) k KKT Inner Iter. κ k T. Test curr. π curr. π new π Overall, average κ k

46 Inexact SQP for nonconvex optimization Outline Motivating Example: Data Assimilation in Weather Forecasting Problem Formulation Solution Techniques Inexact SQP Methods for Equality Constrained Optimization Background and Basics Algorithm Overview Numerical Results Optimization Test Problems: Robustness PDE-Constrained Test Problems: Practicality Related and Future Work Inexact SQP for nonconvex optimization

47 Inexact SQP for nonconvex optimization Convexity assumption Step computation procedure min d R n f k + g T k d d T W k d s.t. c k + A k d = 0» Wk A T k A k 0»» dk gk + A T k λ = k δ k c k

48 Inexact SQP for nonconvex optimization Nonconvex optimization If W k is not convex in the null space of the constraints No guarantee of descent Step may be unbounded Without a factorization of the primal-dual matrix» Wk A T k A k 0 we may not know if the problem is convex or not We could always set W k to a simple positive definite matrix, but then we may be distorting second order information (if available)

49 Inexact SQP for nonconvex optimization First step: recognizing good steps If the problem is convex, then a solution to»»» Wk A T k dk gk + A T k λ = k A k 0 δ k c k (1) is a good step If the problem is nonconvex, then a solution to (1) may still be a good step Idea: characterize d k based on properties of the decomposition d k = u k + v k (Note: d k 2 = u k 2 + v k 2 ) where A k u k = 0 and v k lies in range(a T k )

50 Inexact SQP for nonconvex optimization Central claim A step is good if either termination test for the inexact SQP algorithm is satisfied and for some small θ > 0 we have Notice that (a) θ u k v k or (b) d T k W k d k θ u k 2 (a) implies that the step d k is sufficiently parallel to v k (b) implies that the curvature is sufficiently positive along d k

51 Inexact SQP for nonconvex optimization Estimating properties of the step Observing we find and v k A k v k / A k = A k d k / A k, θ d k A k d k / A k θ u k v k d T k W k d k θ( d k 2 A k d k 2 / A k 2 ) d T k W k d k θ u k 2

52 Inexact SQP for nonconvex optimization Proposed Algorithm Apply an iterative solver to the primal-dual system Stop if a termination test is satisfied and... step is sufficiently parallel to vk θ d k A k d k / A k... or curvature is sufficiently positive d T k W k d k θ ( d k 2 A k d k 2 / A k 2) If θ d k > A k d k / A k, then set W k W k to satisfy θ d k 2 A k d k 2 / A k 2 dk T W k d k and iterate on perturbed system

53 Conclusion We have: 1. Described an important and challenging sample problem in PDE-constrained optimization 2. Described a globally convergent algorithm in the framework of a powerful algorithm: sequential quadratic programming 3. The algorithm has shown to be robust and efficient for realistic applications 4. Described some ideas for extending the approach to nonconvex problems

Inexact Newton Methods and Nonlinear Constrained Optimization

Inexact Newton Methods and Nonlinear Constrained Optimization Frank E. Curtis EPSRC Symposium Capstone Conference Warwick Mathematics Institute July 2, 2009 Outline PDE-Constrained Optimization Newton