An Inexact Newton Method for Optimization

Size: px

Start display at page:

Download "An Inexact Newton Method for Optimization"

Kimberly Chandler
5 years ago
Views:

1 New York University Brown Applied Mathematics Seminar, February 10, 2009

2 Brief biography New York State College of William and Mary (B.S.) Northwestern University (M.S. & Ph.D.) Courant Institute (Postdoc) Research: Nonlinear Optimization, algorithms and theory Large-scale optimization (PDE-constrained problems, today s talk) Methods with fast detection of infeasibility (MINLP problems) Global convergence mechanisms

3 Outline Motivational Example Algorithm development and theoretical results Experimental results Conclusion and final remarks

4 Outline Motivational Example Algorithm development and theoretical results Experimental results Conclusion and final remarks

radio wave adsorption Results in the killing of tumor cells and

5 Hyperthermia treatment Regional hyperthermia is a cancer therapy that aims at heating large and deeply seated tumors by means of radio wave adsorption Results in the killing of tumor cells and makes them more susceptible to other accompanying therapies; e.g., chemotherapy

for numerical optimization The goal is to heat the tumor to the

6 Hyperthermia treatment planning Computer modeling can be used to help plan the therapy for each patient, and it opens the door for numerical optimization The goal is to heat the tumor to the target temperature of 43 C while minimizing damage to nearby cells

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

8 Large-scale optimization Consider 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 The best contemporary methods are limited by problem size; e.g., sequential quadratic programming (small to moderate sizes) interior-point methods (moderate to large sizes) We want the fast solution of problems with millions of variables

9 Challenges in large-scale optimization Computational issues: Large matrices may not be stored Large matrices may not be factored Algorithmic issues: The problem may be nonconvex The problem may be ill-conditioned Computational/Algorithmic issues: No matrix factorizations makes difficulties more difficult

10 Main contributions ALGORITHMS: Inexact Newton methods for constrained optimization, broadening the potential application of fast optimization algorithms THEORY: Global convergence and the potential for fast local convergence SOFTWARE: new release of Ipopt (Wächter) with Pardiso (Schenk) ARTICLES: An Inexact SQP Method for Equality Constrained Optimization, SIAM Journal on Optimization, 19(1): , with R. H. Byrd and J. Nocedal An Inexact Newton Method for Nonconvex Equality Constrained Optimization, Mathematical Programming, Series A, to appear, with R. H. Byrd and J. Nocedal A Matrix-free Algorithm for Equality Constrained Optimization Problems with Rank-Deficient Jacobians, SIAM Journal on Optimization, to appear, with J. Nocedal and A. Wächter An Interior-Point Algorithm for Large-Scale Nonlinear Optimization with Inexact Step Computations, submitted to SIAM Journal on Scientific Computing, with O. Schenk and A. Wächter

11 Outline Motivational Example Algorithm development and theoretical results Experimental results Conclusion and final remarks

12 Equality constrained optimization Consider The Lagrangian is min f (x) x R n s.t. c(x) = 0 L(x, λ) f (x) + λ T c(x) so the first-order optimality conditions are [ ] f (x) + c(x)λ L(x, λ) = F(x, λ) = 0 c(x)

13 Inexact Newton methods Solve F(x, λ) = 0 or min ϕ(x, λ) 1 F(x, λ) 2 2 Inexact Newton methods compute F(x k, λ k )d k = F(x k, λ k ) + r k requiring (Dembo, Eisenstat, Steihaug (1982)) r k κ F(x k, λ k ), κ (0, 1)

14 A naïve Newton method for optimization 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

15 A naïve Newton method for optimization 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

16 A naïve Newton method for optimization fails easily Consider the problem min f (x) = x 1 + x 2, s.t. c(x) = x x = 0

17 Merit function Simply minimizing [ ] ϕ(x, λ) = 1 F(x, 2 λ) 2 = 1 f (x) + c(x)λ 2 2 c(x) is generally inappropriate for optimization We use the merit function φ(x; π) f (x) + π c(x) where π is a penalty parameter

18 Algorithm 0: Newton method for optimization (Assume the problem is convex and regular) for k = 0, 1, 2,... Solve the primal-dual (Newton) equations [ ] [ ] [ ] H(xk, λ k ) c(x k ) dk f (xk ) + c(x c(x k ) T = k )λ k 0 c(x k ) δ k Increase π, if necessary, so that π k λ k + δ k (yields Dφ k (d k ; π k ) 0) Backtrack from α k 1 to satisfy the Armijo condition 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 )

19 Newton methods and sequential quadratic programming If H(x k, λ k ) is positive definite on the null space of c(x k ) T, then [ ] [ ] [ ] H(xk, λ k ) c(x k ) d f (xk ) + c(x c(x k ) T = k )λ k 0 δ c(x k ) is equivalent to min f (x k) + f (x k ) T d + 1 d T H(x d R n 2 k, λ k )d s.t. c(x k ) + c(x k ) T d = 0

20 Minimizing a penalty function Consider the 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

21 Convergence of Algorithm 0 Assumption The sequence {(x k, λ k )} is contained in a convex set Ω over which f, c, and their first derivatives are bounded and Lipschitz continuous. Also, (Regularity) c(x k ) T has full row rank with singular values bounded below by a positive constant (Convexity) u T H(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 (Han (1977)) The sequence {(x k, λ k )} yields the limit [ ] lim f (xk ) + c(x k )λ k k = 0 c(x k )

22 Incorporating inexactness Iterative as opposed to direct methods Compute [ ] [ ] [ ] H(xk, λ k ) c(x k ) dk f (xk ) + c(x c(x k ) T = k )λ k + 0 c(x k ) δ k [ ρk r k ] satisfying [ ρk r k ] [ ] κ f (xk ) + c(x k )λ k, κ (0, 1) c(x k ) If κ is not sufficiently small (e.g., 10 3 vs ), then d k may be an ascent direction for our merit function; i.e., Dφ k (d k ; π k ) > 0 for all π k π k 1

23 Model reductions Define the model of φ(x; π): m(d; π) f (x) + f (x) T d + π( c(x) + c(x) T d ) d k is acceptable if m(d k ; π k ) m(0; π k ) m(d k ; π k ) = f (x k ) T d k + π k ( c(x k ) c(x k ) + c(x k ) T d k ) 0 This ensures Dφ k (d k ; π k ) 0 (and more)

24 Termination test 1 The search direction (d k, δ k ) is acceptable if [ ] [ ] ρk κ f (xk ) + c(x k )λ k, κ (0, 1) c(x k ) r k and if for π k = π k 1 and some σ (0, 1) we have m(d k ; π k ) max{ 1 2 d T k H(x k, λ k )d k, 0} + σπ k max{ c(x k ), r k c(x k ) } }{{} 0 for any d

25 Termination test 2 The search direction (d k, δ k ) is acceptable if ρ k β c(x k ), β > 0 and r k ɛ c(x k ), ɛ (0, 1) Increasing the penalty parameter π then yields m(d k ; π k ) max{ 1 2 d T k H(x k, λ k )d k, 0} + σπ k c(x k ) }{{} 0 for any d

26 Algorithm 1: Inexact Newton for optimization (Byrd, Curtis, Nocedal (2008)) for k = 0, 1, 2,... Iteratively solve [ ] [ ] [ ] H(xk, 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 π so { π k max π k 1, f (x k) T d k + max{ 1 d } T 2 k H(x k, λ k )d k, 0} (1 τ)( c(x k ) r k ) Backtrack from α k 1 to satisfy δ k 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 )

27 Convergence of Algorithm 1 Assumption The sequence {(x k, λ k )} is contained in a convex set Ω over which f, c, and their first derivatives are bounded and Lipschitz continuous. Also, (Regularity) c(x k ) T has full row rank with singular values bounded below by a positive constant (Convexity) u T H(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 )

28 Handling nonconvexity and rank deficiency There are two assumptions we aim to drop: (Regularity) c(xk ) T has full row rank with singular values bounded below by a positive constant (Convexity) u T H(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 e.g., the problem is not regular if it is infeasible, and it is not convex if there are maximizers and/or saddle points Without them, Algorithm 1 may stall or may not be well-defined

29 No factorizations means no clue We might not store or factor [ H(xk, λ k ) ] c(x k ) c(x k ) T 0 so we might not know if the problem is nonconvex or ill-conditioned Common practice is to perturb the matrix to be [ ] H(xk, 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 Poor choices of ξ 1 and ξ 2 can have terrible consequences in the algorithm

30 Our approach for global convergence Decompose the direction d k into a normal component (toward the constraints) and a tangential component (toward optimality) Without convexity, we do not guarantee a minimizer, but our merit function biases the method to avoid maximizers and saddle points

31 Normal component computation (Approximately) solve 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 ) We only require Cauchy decrease: c(x k ) c(x k ) + c(x k ) T v k ɛ v ( c(x k ) c(x k ) + α 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

32 Tangential component computation (idea #1) Standard practice is to then (approximately) solve min ( f (x k ) + H(x k, λ k )v k ) T u ut H(x k, λ k )u s.t. c(x k ) T u = 0, u k However, maintaining can be expensive c(x k ) T u 0 and u k

33 Tangential component computation Instead, we formulate the primal-dual system [ ] [ ] H(xk, λ k ) c(x k ) uk c(x k ) T 0 δ k [ ] f (xk ) + c(x = k )λ k + H(x k, λ k )v k 0 Our ideas from before apply!

34 Handling nonconvexity Convexify the Hessian as in [ H(xk, λ k ) + ξ 1I ] c(x k ) c(x k ) T 0 by monitoring iterates Hessian modification strategy: Increase ξ 1 whenever u k 2 > ψ v k 2, ψ > ut k (H(x k, λ k ) + ξ 1I )u k < θ u k 2, θ > 0

35 Inexact Newton Algorithm 2 (Curtis, Nocedal, Wächter (2009)) for k = 0, 1, 2,... Approximately solve min 1 2 c(x k) + c(x k ) T v 2, s.t. v ω ( c(x k ))c(x k ) to compute v k satisfy Cauchy decrease Iteratively solve [ ] [ ] [ ] H(xk, λ k ) + ξ 1 I 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, increasing ξ 1 as described If only termination test 2 is satisfied, increase π so { π k max π k 1, f (x k) T d k + max{ 1 2 ut k (H(x } k, λ k ) + ξ 1 I )u k, θ u k 2 } (1 τ)( c(x k ) c(x k ) + c(x k ) T d k ) Backtrack from α k 1 to satisfy 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 )

36 Convergence of Algorithm 2 Assumption The sequence {(x k, λ k )} is contained in a convex set Ω over 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

37 Handling inequalities Interior point methods are attractive for large applications Line-search interior point methods that enforce c(x k ) + c(x k ) T d k = 0 may fail to converge globally (Wächter, Biegler (2000)) Fortunately, the trust region subproblem we use to regularize the constraints also saves us from this type of failure!

38 Algorithm 2 (Interior-point version) Apply Algorithm 2 to the logarithmic-barrier subproblem for µ 0 Define min f (x) µ q ln s i, s.t. c E (x) = 0, c I (x) s = 0 i=1 H(x k, λ E,k, λ I,k ) 0 c E (x k ) c I (x k ) d x k 0 µi 0 S k c E (x k ) T d s k δ E,k c I (x k ) T S k 0 0 δ I,k so that the iterate update has [ ] [ ] [ ] xk+1 xk d x + α k s k+1 s k k S k dk s 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)}

39 Convergence of Algorithm 2 (Interior-point) Assumption The sequence {(x k, λ E,k, λ I,k )} is contained in a convex set Ω over which f, c E, c I, and their first derivatives are bounded and Lipschitz continuous Theorem (Curtis, Schenk, Wächter (2009)) For a given µ, Algorithm 2 yields the same limits as in the equality constrained case 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 exist Lagrange multipliers λ such that the first-order optimality conditions of the nonlinear program are satisfied

40 Outline Motivational Example Algorithm development and theoretical results Experimental results Conclusion and final remarks

41 Implementation details Incorporated in IPOPT software package (Wächter) Linear systems solved with PARDISO (Schenk) Symmetric quasi-minimum residual method (Freund (1994)) PDE-constrained model problems 3D grid Ω = [0, 1] [0, 1] [0, 1] Equidistant Cartesian grid with N grid points 7-point stencil for discretization

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

43 Hyperthermia Treatment Planning { min 1 37 in 2 Ω (y(x) Ω\Ω0 yt(x))2 dx, // y t(x) = 43 in Ω 0 u j = a j e iφ j s.t. y(x) 10(y(x) 37) = u M(x)u, in Ω // M jk (x) =< E j (x), E k (x) > E j = sin(jx 1 x 2 x 3 π) 37.0 y(x) 37.5, on Ω 42.0 y(x) 44.0, in Ω 0, // Ω 0 = [3/8, 5/8] 3 N n p q # nnz f # iter CPU sec (direct)

44 Sample solution for N = 40

of Basel) Hyperthermia treatment planning with real patient

45 Numerical experiments (currently underway) Joint with Andreas Wächter (IBM) and Olaf Schenk (U. of Basel) Hyperthermia treatment planning with real patient geometry (with Matthias Christen, U. of Basel) Image registration (with Stefan Heldmann, U. of Lübeck)

46 Outline Motivational Example Algorithm development and theoretical results Experimental results Conclusion and final remarks

47 Conclusion and final remarks Inexact Newton method for optimization with theoretical foundation Convergence guarantees are as good as exact methods, sometimes better Numerical experiments are promising so far, and more to come

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