An Inexact Newton Method for Nonlinear Constrained Optimization

Size: px

Start display at page:

Download "An Inexact Newton Method for Nonlinear Constrained Optimization"

Pauline Eaton
5 years ago
Views:

1 An Inexact Newton Method for Nonlinear Constrained Optimization Frank E. Curtis Numerical Analysis Seminar, January 23, 2009

2 Outline Motivation and background Algorithm development and theoretical results Experimental results Conclusion and final remarks

3 Outline Motivation and background Algorithm development and theoretical results Experimental results Conclusion and final remarks

4 Very large-scale optimization Consider a constrained optimization problem (NLP) of the form min f (x) x R n s.t. c E(x) = 0 c I(x) 0 where f : R n R, c E : R n R p and c I : R n R q are smooth functions We are interested in problems for which the best contemporary methods, i.e., penalty methods sequential quadratic programming interior-point methods cannot be employed due to problem size

Hyperthermia treatment planning Regional hyperthermia is a cancer therapy that aims at heating large and deeply seated tumors by means of radio wave adsorption,

5 Hyperthermia treatment planning Regional hyperthermia is a cancer therapy that aims at heating large and deeply seated tumors by means of radio wave adsorption, which results in the killing of tumor cells and makes them more susceptible to accompanying radio or chemotherapy. See

6 Hyperthermia treatment planning as a NLP The problem can be formulated as { 37 in Ω\Ω0 min y(x) y y,u t(x)dx where y t(x) = 43 in Ω 0 Ω subject to the bio-heat transfer equation 1 (κ(x) y(x)) }{{} thermal conductivity and the bound constraints where Ω 0 is the tumor domain + ω(x, y(x))π(x)(y(x) y b ) = }{{} effects of blood flow 37.0 y(x) 37.5, on Ω 41.0 y(x) 45.0, in Ω 0 σ 2 i u ie i 2 }{{} electromagnetic field, in Ω 1 Pennes (1948)

7 Newton s method An approach that sounds ideal for, say, the equality constrained problem min f (x) x R n s.t. c(x) = 0 (1.1) would be a Newton method The Lagrangian for (1.1) is given by L(x, λ) f (x) + λ T c(x) so if f and c are differentiable, the first-order optimality conditions are [ ] f (x) + c(x)λ L(x, λ) = = 0. c(x) (A single system of nonlinear equalities.)

8 Inexact Newton methods A Newton method for the nonlinear system of equations F(x) = 0 has the iteration ( F(x k ))d k = F(x k ) Applying an iterative linear system solver to this system yields ( F(x k ))d k = F(x k ) + r k and if progress is judged by the merit function φ(x) 1 2 F(x) 2 then the (inner) iteration may be terminated as soon as r k κ k F(x k ) (for superlinear convergence, choose κ k 0) 2 2 Dembo, Eisenstat, Steihaug (1982)

9 A naïve Newton method for NLP Consider the problem min f (x) = x 1 + x 2, s.t. c(x) = x x = 0 that has the first-order optimality conditions F(x, λ) = 1 + 2x 1λ 1 + 2x 2λ = 0 x1 2 + x2 2 1 A Newton method applied to this problem yields 1 k F(x 2 k, λ k ) e e e e e-15

10 A naïve Newton method for NLP Consider the problem min f (x) = x 1 + x 2, s.t. c(x) = x x = 0 that has the first-order optimality conditions F(x, λ) = 1 + 2x 1λ 1 + 2x 2λ = 0 x1 2 + x2 2 1 A Newton method applied to this problem yields 1 k F(x 2 k, λ k ) e e e e e-15 k f (x k ) c(x k ) e e e e e e e e e e-08

11 Our goal Our goal is to design an inexact Newton method for constrained optimization Such a framework is important so the user will be able to balance the computational costs between outer (optimization) and inner (step computation) iterations for their own applications The method must always remember that an optimization problem is being solved Many of the usual techniques in Newton-like methods for optimization for handling non-convex objective and constraint functions (near) rank deficiency of the constraint Jacobian inequality constraints are not applicable here! (We need to get creative.)

12 Outline Motivation and background Algorithm development and theoretical results Experimental results Conclusion and final remarks

13 A merit function for optimization The failure in applying a standard (inexact) Newton method to an optimization problem is due to the choice of merit function! That is, [ ] φ(x, λ) = 1 F(x, 2 λ) 2 = 1 f (x) + c(x)λ 2 2 c(x) is inappropriate for purposes of optimization Our method will be centered around the merit function φ(x; π) f (x) + π c(x) (known as an exact penalty function)

14 Algorithm 0: Newton method for optimization for k = 0, 1, 2,... Evaluate f (x k ), f (x k ), c(x k ), c(x k ), and 2 xxl(x k, λ k ) Solve the primal-dual equations [ ] [ ] [ ] 2 xx L(x k, λ k ) c(x k ) dk f (xk ) + c(x c(x k ) T = k )λ k 0 c(x k ) Increase the penalty parameter, if necessary, so that Dφ k (d k ; π k ) 0 Perform a line search to find α k (0, 1] satisfying the Armijo condition δ k φ(x k + α k d k ; π k ) φ(x k ; π k ) + ηα k Dφ k (d k ; π k ) Update iterate (x k+1, λ k+1 ) (x k, λ k ) + α k (d k, δ k )

15 Newton methods and sequential quadratic programming The primal-dual system [ ] [ ] 2 xx L(x, λ) c(x) d c(x) T 0 δ is equivalent to the SQP subproblem [ ] f (x) + c(x)λ = c(x) min d R n f (x) + f (x)t d d T ( 2 xxl(x, λ))d s.t. c(x) + c(x) T d = 0 if 2 xxl(x, λ) is positive definite on the null space of c(x) T

16 Minimizing an exact penalty function Consider the exact penalty function for min (x 1) 2, s.t. x = 0 i.e. φ(x; π) = (x 1) 2 + π x for different values of the penalty parameter π Figure: π = 1 Figure: π = 2

17 Convergence of Algorithm 0 Assumption The sequence {(x k, λ k )} is contained in a convex set Ω and f, c, and their first derivatives are bounded and Lipschitz continuous on Ω (Regularity) c(x k ) T has full row rank with smallest singular value bounded below by a positive constant (Convexity) u T ( 2 xxl(x k, λ k ))u µ u 2 for µ > 0 for all u R n satisfying u 0 and c(x k ) T u = 0 Theorem The sequence {(x k, λ k )} yields the limit [ ] lim f (xk ) + c(x k )λ k k = 0 c(x k )

18 Incorporating inexactness An inexact solution to the primal-dual equations; i.e., [ ] [ ] [ 2 xx L(x, λ) c(x) d f (x) + c(x)λ c(x) T = 0 δ c(x) ] + [ ] ρ r may yield Dφ k (d k ; π k ) > 0 for any π k π k 1, even if for some κ (0, 1) we have [ ] [ ] ρ κ f (x) + c(x)λ r c(x) (the standard condition required in inexact Newton methods) It is the merit function that tells us what steps are appropriate, and so it should be what tells us which inexact solutions to the Newton (primal-dual) system are acceptable search directions

19 Model reductions Modern optimization techniques focus on models of the problem functions We apply an iterative solver to the primal-dual system [ ] [ ] [ ] 2 xx L(x, λ) c(x) d f (x) + c(x)λ c(x) T = 0 δ c(x) (since an exact solution will produce an acceptable search direction) A search direction is deemed acceptable if the improvement in the model m(d; π) f (x) + f (x) T d + π( c(x) + c(x) T d ) is sufficiently large; i.e., if we have m(d; π) m(0; π) m(d; π) = f (x) T d + π( c(x) c(x) + c(x) T d ) 0

20 Termination test 1 An inexact solution (d, δ) to the primal-dual system is acceptable if [ ] [ ] ρ κ f (x) + c(x)λ r c(x) for some κ (0, 1) and if m(d; π) max{ 1 d T ( 2 2 xxl(x, λ))d, 0} + σπ max{ c(x), r c(x) } for some σ (0, 1)

21 Termination test 2 An inexact solution to the primal-dual system is acceptable if for some β > 0 and ɛ (0, 1) ρ β c(x) r ɛ c(x) Increasing the penalty parameter can then yield m(d; π) max{ 1 2 d T ( 2 xxl(x, λ))d, 0} + σπ c(x)

22 Algorithm 1: Inexact Newton for optimization for k = 0, 1, 2,... Evaluate f (x k ), f (x k ), c(x k ), c(x k ), and 2 xxl(x k, λ k ) Iteratively solve the primal-dual equations [ ] [ ] [ ] 2 xx L(x k, λ k ) c(x k ) dk f (xk ) + c(x c(x k ) T = k )λ k 0 c(x k ) until termination test 1 or 2 is satisfied If only termination test 2 is satisfied, increase the penalty parameter so δ k π k f (x k) T d k + max{ 1 2 d T k ( 2 xxl(x k, λ k ))d k, 0} (1 τ)( c(x k ) r k ) Perform a line search to find α k (0, 1] satisfying the sufficient decrease condition φ(x k + α k d k ; π k ) φ(x k ; π k ) ηα k m(d k ; π k ) Update iterate (x k+1, λ k+1 ) (x k, λ k ) + α k (d k, δ k )

23 Convergence of Algorithm 1 Assumption The sequence {(x k, λ k )} is contained in a convex set Ω and f, c, and their first derivatives are bounded and Lipschitz continuous on Ω (Regularity) c(x k ) T has full row rank with smallest singular value bounded below by a positive constant (Convexity) u T ( 2 xxl(x k, λ k ))u µ u 2 for µ > 0 for all u R n satisfying u 0 and c(x k ) T u = 0 Theorem (Byrd, Curtis, Nocedal (2008)) The sequence {(x k, λ k )} yields the limit [ ] lim f (xk ) + c(x k )λ k k = 0 c(x k )

24 Take-away idea #1 In a (inexact) Newton method for optimization, global convergence depends on the use of a merit function appropriate for optimization. This function should dictate which inexact solutions to the primal-dual system are acceptable search directions.

25 Handling nonconvexity and rank deficiency There are two assumptions we aim to drop: (Regularity) c(xk ) T has full row rank with smallest singular value bounded below by a positive constant (Convexity) u T ( 2 xxl(x k, λ k ))u µ u 2 for µ > 0 for all u R n satisfying u 0 and c(x k ) T u = 0 If the constraints are ill-conditioned, Algorithm 1 may compute long unproductive search directions If the problem is non-convex, Algorithm 1 can easily converge to a maximizer or saddle point There is also a danger in each case that the algorithm may stall during a given iteration since an acceptable search direction cannot be computed

26 No factorizations means no clue A significant challenge in overcoming these obstacles is that we have no way of recognizing when the problem is non-convex or ill-conditioned, since we do not form or factor the primal-dual matrix [ ] 2 xx L(x k, λ k ) c(x k ) c(x k ) T 0 Common practice is to perturb the matrix to be [ ] 2 xx L(x k, λ k ) + ξ 1I c(x k ) c(x k ) T ξ 2I where ξ 1 convexifies the model and ξ 2 regularizes the constraints, but arbitrary choices of these parameters can have terrible consequences on the behavior of the algorithm

27 Decomposing the step into two components We guide the step computation by decomposing the primal search direction into a normal component (toward satisfying the constraints) and a tangential component (toward optimality)

28 Normal component computation We define the trust region subproblem for some ω > 0 min 1 2 c(x k) + c(x k ) T v 2 s.t. v ω ( c(x k ))c(x k ) An (approximate) solution v k satisfying the Cauchy decrease condition c(x k ) c(x k ) + c(x k ) T v k ɛ v ( c(x k ) c(x k ) + c(x k ) T ṽ k ) for ɛ v (0, 1), where ṽ k = ( c(x k ))c(x k ) is the direction of steepest descent, can be computed without any matrix factorizations (conjugate gradient method, inexact dogleg method)

29 Normal component computation We define the trust region subproblem for some ω > 0 min 1 2 c(x k) + c(x k ) T v 2 s.t. v ω ( c(x k ))c(x k ) An (approximate) solution v k satisfying the Cauchy decrease condition c(x k ) c(x k ) + c(x k ) T v k ɛ v ( c(x k ) c(x k ) + c(x k ) T ṽ k ) for ɛ v (0, 1), where ṽ k = ( c(x k ))c(x k ) is the direction of steepest descent, can be computed without any matrix factorizations (conjugate gradient method, inexact dogleg method)

30 Tangential component computation (idea #1) Standard practice is to then consider the trust region subproblem min ( f (x k ) + 2 xxl(x k, λ k )v k ) T u ut ( 2 xxl(x k, λ k ))u s.t. c(x k ) T u = 0, u k Note that an exact solution would yield c(x k ) T d k = c(x k ) T v k An iterative procedure for solving this problem (e.g., a projected conjugate gradient method), however, requires repeated (inexact) projections of vectors onto the null space of c(x k ) T (to maintain c(x k ) T u 0 while respecting the trust region constraint)

31 Tangential component computation Instead, we formulate the primal-dual system [ ] [ ] 2 xx L(x k, λ k ) c(x k ) uk c(x k ) T 0 δ k [ ] f (xk ) + c(x = k )λ k + xxl(x 2 k, λ k )v k 0 Our ideas for assessing search directions based on reductions in a model of an exact penalty function carry over to this perturbed system

32 Handling non-convexity The only remaining component is a mechanism for handling non-convexity We follow the idea of convexifying the Hessian matrix as in [ 2 xx L(x k, λ k ) + ξ 1I ] c(x k ) c(x k ) T 0 by monitoring properties of the computed trial search directions Hessian modification strategy: During the iterative step computation, modify the Hessian matrix by increasing its smallest eigenvalue whenever an approximate solution satisfies for ψ, θ > 0 u k > ψ v k 1 2 ut k ( xxl(x 2 k, λ k ))u k < θ u k 2

33 Inexact Newton Algorithm 2 for k = 0, 1, 2,... Evaluate f (x k ), f (x k ), c(x k ), c(x k ), and 2 xx L(x k, λ k ) Compute an approximate solution v k to the trust region subproblem min 1 2 c(x k) + c(x k ) T v 2, s.t. v ( c(x k ))c(x k ) satisfying the normal component condition Iteratively solve the primal-dual equations [ ] [ ] [ ] 2 xx L(x k, λ k ) c(x k ) dk f (xk ) + c(x c(x k ) T = k )λ k 0 δ k c(x k ) T v k until termination test 1 or 2 is satisfied, modifying 2 xx L(x k, λ k ) based on the Hessian modification strategy If only termination test 2 is satisfied, increase the penalty parameter so π k f (x k) T d k + max{ 1 2 ut k ( 2 xx L(x k, λ k ))u k, θ u k 2 } (1 τ)( c(x k ) c(x k ) + c(x k ) T d k ) Perform a line search to find α k (0, 1] satisfying the sufficient decrease condition φ(x k + α k d k ; π k ) φ(x k ; π k ) ηα k m(d k ; π k ) Update iterate (x k+1, λ k+1 ) (x k, λ k ) + α k (d k, δ k )

34 Convergence of Algorithm 2 Assumption The sequence {(x k, λ k )} is contained in a convex set Ω on which f, c, and their first derivatives are bounded and Lipschitz continuous Theorem (Curtis, Nocedal, Wächter (2009)) If all limit points of { c(x k ) T } have full row rank, then the sequence {(x k, λ k )} yields the limit [ ] lim f (xk ) + c(x k )λ k k = 0. c(x k ) Otherwise, and if {π k } is bounded, then lim ( c(x k))c(x k ) = 0 k lim f (x k) + c(x k )λ k k

35 Take-away ideas #2 and #3 In an inexact Newton method for optimization... ill-conditioning can be handled by regularizing the constraint model with a trust region. However, a trust region for the tangential component may result in an expensive algorithm. The method we propose avoids these costs and emulates a true inexact Newton method. non-convexity can be handled by monitoring properties of the iterative solver iterates and modifying the Hessian only when it appears that a sufficient model reduction may not be obtained. This process may not (and need not) result in a convex model.

36 Handling inequalities The most attractive class of algorithms for our purposes are interior point methods A class of interior point methods that satisfying c(x k ) + c(x k ) T d k = 0 may fail to converge from remote starting points 3 Fortunately, the trust region subproblem we use to regularize the constraints also saves us from this type of failure! 3 (Wächter, Biegler (2000)

37 Algorithm 2 (Interior-point version) We apply Algorithm 2 to the logarithmic-barrier subproblem q min f (x) µ ln s i, s.t. c E(x) = 0, c I(x) = s for µ 0 i=1 We scale quantities related to the slack variables d s k so that the step computation subproblems still have the form min 1 2 c(x k, s k ) + A(x k, s k )v 2, δ k s.t. v ω A(x k, s k ) T c(x k, s k ) and [H(xk ] [ ] [ ], s k, λ k ) A(x k, s k ) T dk g(xk, s = k ) + A(x k, s k ) T λ k A(x k, s k ) 0 c(x k, s k ) where the Hessian H k and constraint Jacobian A k have the same properties as before We incorporate a fraction-to-the-boundary rule in the line search and a slack reset in the algorithm to maintain s max{0, c I(x)}

38 Convergence of Algorithm 2 (Interior-point) Under the same assumption as before, our convergence theorem for Algorithm 2 still holds for the barrier subproblem for a given µ Theorem (Curtis, Schenk, Wächter (2009)) If Algorithm 2 yields a sufficiently accurate solution to the barrier subproblem for each {µ j } 0 and if the linear independence constraint qualification (LICQ) holds at a limit point x of {x j }, then there exists Lagrange multipliers λ such that the first-order optimality conditions of the nonlinear program are satisfied

39 Take-away ideas #4 In an inexact Newton method for optimization... inequality constraints can be easily incorporated into Algorithm 2 if quantities related to the slack variables are scaled appropriately

40 Outline Motivation and background Algorithm development and theoretical results Experimental results Conclusion and final remarks

41 Implementation details We implemented our interior-point version of Algorithm 2 in the IPOPT software package and used PARDISO for the iterative linear system solves The normal and tangential (primal-dual) step computation was performed with the symmetric quasi-minimum residual method (SQMR) We solved two model PDE-constrained problems on the three-dimensional grid Ω = [0, 1] [0, 1] [0, 1], using an equidistant Cartesian grid with N grid points in each spatial direction and a standard 7-point stencil for discretizing the operators

42 Boundary control problem Let u(x) be defined on Ω and solve min 1 2 Ω (y(x) yt(x))2 dx s.t. (e y(x) y(x)) = 20, y(x) = u(x), on Ω 2.5 u(x) 3.5, on Ω in Ω where y t(x) = x 1 (x 1 1)x 2 (x 2 1) sin(2πx 3 ) N n p q # nnz f # iter CPU sec e e e e e (direct) e

43 (Simplified) Hyperthermia Treatment Planning Let u j = a j e iφ j be a complex vector of amplitudes a R 10 and phases φ R 10 of 10 antennas, let M ij (x) =< E i (x), E j (x) > with E j = sin(jx 1 x 2 x 3 π), and let the tumor be defined by Ω 0 = [3/8, 5/8] 3, and solve min 1 2 Ω (y(x) yt(x))2 dx s.t. y(x) 10(y(x) 37) = u M(x)u, in Ω 37.0 y(x) 37.5, on Ω 42.0 y(x) 44.0, in Ω 0 where { 37 in Ω\Ω0 y t(x) = 43 in Ω 0 N n p q # nnz f # iter CPU sec (direct)

44 (Simplified) Hyperthermia Treatment Planning An example solution (N = 40) Figure: y(x) Figure: y t (x)

45 Numerical experiments (coming soon) Further numerical experimentation is now underway with Andreas Wächter (IBM) and Olaf Schenk (U. of Basel) Hyperthermia treatment planning with real patient geometry (with Matthias Christen, U. of Basel) Geophysical modeling applications (with Johannes Huber, U. of Basel) Image registration (with Stefan Heldmann, U. of Lübeck)

46 Outline Motivation and background Algorithm development and theoretical results Experimental results Conclusion and final remarks

47 Conclusion and final remarks We have presented an inexact Newton method for constrained optimization The method hinges on a merit function appropriate for optimization, and the iterative linear system solves focus on a model of this function to decide when to terminate We have extended the basic algorithm to solve non-convex and ill-conditioned problems, and to solve problems with inequality constraints The algorithm is globally convergent to first-order optimal points or infeasible stationary points Numerical experiments are promising so far, and further testing on real-world problems is underway

An Inexact Newton Method for Optimization

An Inexact Newton Method for Optimization New York University Brown Applied Mathematics Seminar, February 10, 2009 Brief biography New York State College of William and Mary (B.S.) Northwestern University (M.S. & Ph.D.) Courant Institute (Postdoc)