Large-Scale Nonlinear Optimization with Inexact Step Computations

Transcription

1 Large-Scale Nonlinear Optimization with Inexact Step Computations Andreas Wächter IBM T.J. Watson Research Center Yorktown Heights, New York IPAM Workshop on Numerical Methods for Continuous Optimization October 14, 2010 (joint with Frank Curtis, Johannes Huber, and Olaf Schenk) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

2 Nonlinear Optimization min x R n f (x) s.t. c(x) = 0 d(x) 0 x f (x) : R n R c(x) : R n R m d(x) : R n R p Variables Objective function Equality constraints Inequality constraints Functions f (x), c(x), d(x) are smooth (C 2 ); sparse derivatives Want to find local solution Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

3 Nonlinear Optimization min x R n f (x) s.t. c(x) = 0 d(x) 0 x f (x) : R n R c(x) : R n R m d(x) : R n R p Variables Objective function Equality constraints Inequality constraints Functions f (x), c(x), d(x) are smooth (C 2 ); sparse derivatives Want to find local solution Barrier method: p min f (x) µ ln(s (i) ) x R n,s R p i=1 s.t. c(x) = 0 d(x) + s = 0 Solve sequence of barrier problems In following, concentrate on NLPs only with equality constraints (fixed µ) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

4 Motivation Example PDE constrained optimization problem: 3D inverse problem [Haber et al.] All-at-once approach Finite difference discretization on regular grid N = 32 in each dimension Size of NLP: 229,376 variables 196,608 equality and 65,536 inequality constraints Using Ipopt with Pardiso linear solver (direct factorization) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

5 Motivation Example PDE constrained optimization problem: 3D inverse problem [Haber et al.] All-at-once approach Finite difference discretization on regular grid N = 32 in each dimension Size of NLP: 229,376 variables 196,608 equality and 65,536 inequality constraints Using Ipopt with Pardiso linear solver (direct factorization) About 15 hours for first iteration! Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

6 Motivation Example PDE constrained optimization problem: 3D inverse problem [Haber et al.] All-at-once approach Finite difference discretization on regular grid N = 32 in each dimension Size of NLP: 229,376 variables 196,608 equality and 65,536 inequality constraints Using Ipopt with Pardiso linear solver (direct factorization) About 15 hours for first iteration! # nonzeros in sparse step computation matrix: 6,329,867 # nonzeros in factor: 2,503,219,440 Fill-in factor: 395 (!) Want to be able to use iterative linear solver Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

7 Motivation Example PDE constrained optimization problem: 3D inverse problem [Haber et al.] All-at-once approach Finite difference discretization on regular grid N = 32 in each dimension Size of NLP: 229,376 variables 196,608 equality and 65,536 inequality constraints Using Ipopt with Pardiso linear solver (direct factorization) About 15 hours for first iteration! # nonzeros in sparse step computation matrix: 6,329,867 # nonzeros in factor: 2,503,219,440 Fill-in factor: 395 (!) Want to be able to use iterative linear solver Looking for other applications with large fill-in (ideas?) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

8 Outline Newton s method for F(x) = 0 with inexact step computation NLP algorithm with inexact step computation Global convergence result Numerical experiments (3D PDE problems) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

9 Newton s Method min x R n NLP f (x) s.t. c(x) = 0 Optimality Conditions f (x) + c(x)y = 0 c(x) = 0 Apply Newton s Method [ ]( ) ( ) Wk c(x k ) xk f (xk ) + c(x c(x k ) T = k )y k 0 y k c(x k ) W k = xx L(x k, y k ) Hessian of Lagrangian (or approximation) L(x, y) = f (x) + y T c(x) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

10 Newton s Method min x R n NLP f (x) s.t. c(x) = 0 Optimality Conditions f (x) + c(x)y = 0 c(x) = 0 Apply Newton s Method [ ]( ) ( ) Wk c(x k ) xk f (xk ) + c(x c(x k ) T = k )y k 0 y k c(x k ) W k = xx L(x k, y k ) Hessian of Lagrangian (or approximation) L(x, y) = f (x) + y T c(x) Simple approach: Consider as root finding problem F(z) = 0 (with z = (x, y)) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

11 Basic Line Search Algorithm for F(z) = 0 Given: Starting point z 0 ; set k 0 1 Solve linear system to compute search direction F(z k ) T z k = F(z k ) 2 Perform backtracking line search to determine step sizeαk (0, 1] satisfying Armijo condition (η (0, 1)): φ(z k +α k z k ) φ(z k ) +ηα k Dφ(z k ; z k ) Here: Merit functionφ(z) = F(z) Update iterates zk+1 = z k +α k z k 4 Increase k k + 1 and go back to Step 1. Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

12 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

13 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

14 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

15 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

16 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 3 Then 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 k φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

17 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 3 Then 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 k φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) k ηα i Dφ(z i ; z i ) i=0 i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

18 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 3 Then 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 k φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) k ηα i Dφ(z i ; z i ) i=0 i=0 2ηᾱ k F(z i ) 2 2 i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

19 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 3 Then k L φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) k ηα i Dφ(z i ; z i ) i=0 i=0 2ηᾱ k F(z i ) 2 2 i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

20 Simple Global Convergence Proof Step computation F(z k ) T z k = F(z k ) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k = 2 F(z k ) 2 2< 0 Proof outline: 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 3 Then k L φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) k ηα i Dφ(z i ; z i ) i=0 4 Therefore: limk F(z k ) = 0 i=0 2ηᾱ k F(z i ) 2 2 i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

21 Simple Global Convergence Proof Inexact step computation [Dembo, Eisenstat, Steihaug (1982)] F(z k ) T z k = F(z k ) + r k ( r k κ F(z k ) 2, κ (0, 1)) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k 2 F(z k ) F(z k ) T r k Proof outline: 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 3 Then k L φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) k ηα i Dφ(z i ; z i ) i=0 4 Therefore: limk F(z k ) = 0 i=0 2ηᾱ k F(z i ) 2 2 i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

22 Simple Global Convergence Proof Inexact step computation [Dembo, Eisenstat, Steihaug (1982)] F(z k ) T z k = F(z k ) + r k ( r k κ F(z k ) 2, κ (0, 1)) Merit functionφ(z) = F(z) 2 2 Dφ(z k ; z k ) = 2F(z k ) T F(z k ) T z k 2(1 κ) F(z k ) 2 2 Proof outline: 1 Show zk are bounded (e.g., if F(z k ) and F(z k ) 1 bounded) 2 Therefore step sizeαk ᾱ>0 bounded away from 0 3 Then k L φ(z k+1) φ(z 0 ) = (φ(z i+1 ) φ(z i )) k ηα i Dφ(z i ; z i ) i=0 4 Therefore: limk F(z k ) = 0 i=0 2ηᾱ k (1 κ) F(z i ) 2 2 i=0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

23 Discussion Root finding algorithm F(z) = 0 can easily incorporate inexactness of linear solver Natural consistency between step computation and merit function Can be used for solving PDEs Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

24 Discussion Root finding algorithm F(z) = 0 can easily incorporate inexactness of linear solver Natural consistency between step computation and merit function Can be used for solving PDEs Requires F(z k ) 1 is uniformly bounded F(z k ) = Violated, e.g, if c(x k ) loses rank [ Wk ] c(x k ) c(x k ) T 0 In barrier approach, this also corresponds to violation of LICQ Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

25 Discussion Root finding algorithm F(z) = 0 can easily incorporate inexactness of linear solver Natural consistency between step computation and merit function Can be used for solving PDEs Requires F(z k ) 1 is uniformly bounded F(z k ) = Violated, e.g, if c(x k ) loses rank [ Wk ] c(x k ) c(x k ) T 0 In barrier approach, this also corresponds to violation of LICQ Maximizers and stationary points also satisfy F(z) = 0 Would like to encourage convergence to minimizers Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

26 Merit Function For Optimization Classical penalty function:φ ν (x) = f (x) +ν c(x) For ν sufficiently large: x local solution of NLP x local minimizer ofφ ν (x) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

27 Merit Function For Optimization Classical penalty function:φ ν (x) = f (x) +ν c(x) For ν sufficiently large: x local solution of NLP Algorithm: x local minimizer ofφ ν (x) 1 Solve linear system to compute search direction [ ]( ) ( ) Wk c(x k ) xk f (xk ) + c(x c(x k ) T = k )y k 0 y k c(x k ) 2 Perform backtracking line search to determine step sizeαk with φ(x k +α k x k ) φ(x k )+ηα k Dφ(x k ; x k ) [Need Dφ(x k ; x k )<0!] 3 Update iterates (xk+1, y k+1 ) = (x k, y k ) +α k ( x k, y k ) 4 Increase k k + 1 and go back to Step 1. Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

28 Ensuring Descent If c(x k ) T x k + c(x k ) = 0 then Dφ νk (x k ; x k ) = f (x k ) T x k ν k c(x k ) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

29 Ensuring Descent If c(x k ) T x k + c(x k ) = 0 then Dφ νk (x k ; x k ) = f (x k ) T x k ν k c(x k ) To guarantee Dφ νk (x k ; x k )<0: Before line search, possibly increase ν Careful: Need to avoidν! Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

30 Ensuring Descent If c(x k ) T x k + c(x k ) = 0 then Dφ νk (x k ; x k ) = f (x k ) T x k ν k c(x k ) To guarantee Dφ νk (x k ; x k )<0: Before line search, possibly increase ν Careful: Need to avoidν! One option: Convexify local model [ ]( ) ( ) Wk +δi c(x k ) xk f (xk ) + c(x c(x k ) T = k )y k 0 y k c(x k ) Make sure that v T (W k +δi )v> 0 for all v with c(x k ) T v = 0 (δ 0) Can be verified by looking at inertia (requires factorization) E.g., in Ipopt : Chooseδ 0 so that inertia is as desired Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

31 Simple Convergence Proof Show first (based on strongish assumptions): Steps x k are bounded Step sizeα k ᾱ>0 bounded away from 0 ν k = ν for large k Dφ ν (x k ; x k ) θ OptCond k (θ>0) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

32 Simple Convergence Proof Show first (based on strongish assumptions): Steps x k are bounded Step sizeα k ᾱ>0 bounded away from 0 ν k = ν for large k Dφ ν (x k ; x k ) θ OptCond k (θ>0) Then, as before, using Armijo condition: L φ ν (x k+1 ) φ ν (x 0 ) k = (φ ν (x i+1 ) φ ν (x i )) i=0 k ηα i Dφ ν (x i ; x i ) k ηᾱθ OptCont i i=0 i=0 Therefore: lim k OptCont k = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

33 Discussion of Convergence Required for this convergence proof: Dφνk (x k ; x k )<0 νk = ν for large k Steps xk are bounded Step sizeαk ᾱ>0 bounded away from 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

34 Discussion of Convergence Required for this convergence proof: Dφνk (x k ; x k )<0 νk = ν for large k Steps xk are bounded Step sizeαk ᾱ>0 bounded away from 0 Standard algorithm/proof fails if Inertia cannot be computed iterative linear solver Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

35 Discussion of Convergence Required for this convergence proof: Dφνk (x k ; x k )<0 νk = ν for large k Steps xk are bounded Step sizeαk ᾱ>0 bounded away from 0 Standard algorithm/proof fails if Inertia cannot be computed iterative linear solver Step does not satisfy linear system exactly iterative linear solver Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

36 Discussion of Convergence Required for this convergence proof: Dφνk (x k ; x k )<0 νk = ν for large k Steps xk are bounded Step sizeαk ᾱ>0 bounded away from 0 Standard algorithm/proof fails if Inertia cannot be computed iterative linear solver Step does not satisfy linear system exactly iterative linear solver Constraint gradients become linearly dependent strong assumption Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

37 Discussion of Convergence Required for this convergence proof: Dφνk (x k ; x k )<0 νk = ν for large k Steps xk are bounded Step sizeαk ᾱ>0 bounded away from 0 Standard algorithm/proof fails if Inertia cannot be computed iterative linear solver Step does not satisfy linear system exactly iterative linear solver Constraint gradients become linearly dependent strong assumption For barrier methods: The fraction-to-the-boundary rule leads toα k 0 fundamental problem for line-search barrier methods [W., Biegler, 2000] Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

38 Goal Devise an algorithm that can work with an iterative linear solver can deal with rank-deficient constraint Jacobian overcomes convergence problems of line-search barrier methods Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

39 Goal Devise an algorithm that can work with an iterative linear solver can deal with rank-deficient constraint Jacobian overcomes convergence problems of line-search barrier methods Main ingredients: Step decomposition helps to bound step (no need to assume full rank of c(xk )) handle nonconvexity Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

40 Goal Devise an algorithm that can work with an iterative linear solver can deal with rank-deficient constraint Jacobian overcomes convergence problems of line-search barrier methods Main ingredients: Step decomposition helps to bound step (no need to assume full rank of c(xk )) handle nonconvexity Termination test of linear solver based on model of merit function consistency between inexact solution and merit function does not ignore optimization nature of problem Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

41 Useful Equivalence [ W k c(x k ) c(x k ) T 0 ]( ) ( ) xk f (xk ) + c(x = k )y k y k c(x k ) 1 min 2 xt k W k x k + f (x k ) T x k s.t. c(x k ) T x k + c(x k ) = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

42 Step Decomposition min 1 2 xt k W k x k + f (x k ) T x k s.t. c(x k ) T x k + c(x k ) = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

43 Step Decomposition min 1 2 xt k W k x k + f (x k ) T x k s.t. c(x k ) T x k + c(x k ) = 0 x k = n k + t k Normal step nk c(x k ) T n k + c(x k ) = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

44 Step Decomposition min 1 2 xt k W k x k + f (x k ) T x k s.t. c(x k ) T x k + c(x k ) = 0 x k = n k + t k Normal step nk Tangential step tk 1 ( T min t 2 tt W k t + Wk T n k + f (x k )) t s.t. c(x k ) T t = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

45 Inexact Steps (Normal Component) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

46 Inexact Steps (Normal Component) Normal component c(x k ) T n k + c(x k ) = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

47 Inexact Steps (Normal Component) Normal component min n c(x k ) T n + c(x k ) 2 2 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

48 Inexact Steps (Normal Component) Normal component min n c(x k ) T n + c(x k ) 2 2 s.t. n 2 ω c(x k )c(x k ) 2 (ω> 0 fixed, usually large) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

49 Inexact Steps (Normal Component) min n c(x k ) T n + c(x k ) 2 2 s.t. n 2 ω c(x k )c(x k ) 2 Trust-region subproblem Controls size of normal component Trust region inactive ifω [σmin ( c(x k ))] 1 c(xk )c(x k ) goes to zero if x k converges to minimizer of c(x) 2 2 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

50 Inexact Steps (Normal Component) min n c(x k ) T n + c(x k ) 2 2 s.t. n 2 ω c(x k )c(x k ) 2 Trust-region subproblem Controls size of normal component Trust region inactive ifω [σmin ( c(x k ))] 1 c(xk )c(x k ) goes to zero if x k converges to minimizer of c(x) 2 2 Inexact: Make at least as much progress as steepest descent step Cauchy step n C Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

51 Inexact Steps (Normal Component) min n c(x k ) T n + c(x k ) 2 2 s.t. n 2 ω c(x k )c(x k ) 2 Trust-region subproblem Controls size of normal component Trust region inactive ifω [σmin ( c(x k ))] 1 c(xk )c(x k ) goes to zero if x k converges to minimizer of c(x) 2 2 Inexact: Make at least as much progress as steepest descent step Cauchy step n C In practice: [ ]( ) ( ) I c(x k ) δ 0 Shortest norm solution: c(x k ) T 0 n N = c(x k ) (solve only approximately) Choose combination of n C and n N satisfying trust region giving best objective Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

52 Inexact Steps (Tangential Component) Tangential component 1 ( T min t 2 tt W k t + Wk T n k + f (x k )) t s.t. c(x k ) T t = 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

53 Inexact Steps (Tangential Component) Tangential component [ ]( ) ( ) W k c(x k ) tk f (xk ) + c(x c(x k ) T = k )y k + W k n k 0 y k 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

54 Inexact Tangential Component [ ]( ) ( ) W k c(x k ) tk f (xk ) + c(x c(x k ) T = k )y k + W k n k 0 y k 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

55 Inexact Tangential Component [ ]( ) ( ) W k c(x k ) tk f (xk ) + c(x c(x k ) T = k )y k + W k n k 0 y k 0 Rewrite using x k = n k + t k Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

56 Inexact Tangential Component [ ]( ) ( ) W k c(x k ) xk f (xk ) + c(x c(x k ) T = k )y k 0 y k c(x k ) T n k Rewrite using x k = n k + t k Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

57 Inexact Tangential Component [ ) ( ) ( ) W k c(x k ) ]( x k f (xk ) + c(x c(x k ) T = k )y k ρ 0 y k c(x k ) T + k n k r k Rewrite using x k = n k + t k ( x k current inexact solution) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

58 Inexact Tangential Component [ ) ( ) ( ) W k c(x k ) ]( x k f (xk ) + c(x c(x k ) T = k )y k ρ 0 y k c(x k ) T + k n k r k Rewrite using x k = n k + t k ( x k current inexact solution) Require: (simplified) 1 Residual ( ) f (xk ) + c(x ρ k κ k )y k c(x k ) T κ (0, 1) n k Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

59 Inexact Tangential Component [ ) ( ) ( ) W k c(x k ) ]( x k f (xk ) + c(x c(x k ) T = k )y k ρ 0 y k c(x k ) T + k n k r k Rewrite using x k = n k + t k ( x k current inexact solution) Require: (simplified) 1 Residual ( ) f (xk ) + c(x ρ k κ k )y k c(x k ) T κ (0, 1) n k 2 Tangential component condition: Satisfy at least one of T t k Wk t k θ t k 2 θ>0 t k ψ n k ψ> 0 Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

60 Inexact Tangential Component [ ) ( ) ( ) W k c(x k ) ]( x k f (xk ) + c(x c(x k ) T = k )y k ρ 0 y k c(x k ) T + k n k r k Rewrite using x k = n k + t k ( x k current inexact solution) Require: (simplified) 1 Residual ( ) f (xk ) + c(x ρ k κ k )y k c(x k ) T κ (0, 1) n k 2 Tangential component condition: Satisfy at least one of T t k Wk t k θ t k 2 θ>0 t k ψ n k ψ> 0 Update W k W k +δi if both violated Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

61 Model Reduction Condition Merit function: φ νk (x) = f (x) +ν k c(x) Linear model: m νk (x k, x) = f (x k ) + f (x k ) T x +ν k c(x k ) + c(x k ) T x Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

62 Model Reduction Condition Merit function: φ νk (x) = f (x) +ν k c(x) Linear model: m νk (x k, x) = f (x k ) + f (x k ) T x +ν k c(x k ) + c(x k ) T x Predicted reduction: m νk (x k ; x) = m νk (x k, 0) m νk (x k, x) Can show Dφ νk (x k ; x) m νk (x k ; x) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

63 Model Reduction Condition Merit function: φ νk (x) = f (x) +ν k c(x) Linear model: m νk (x k, x) = f (x k ) + f (x k ) T x +ν k c(x k ) + c(x k ) T x Predicted reduction: m νk (x k ; x) = m νk (x k, 0) m νk (x k, x) Can show Dφ νk (x k ; x) m νk (x k ; x) Require: m νk (x k ; x k ) is sufficiently positive, (= guarantee direction of descent for currentν k ) or Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

64 Model Reduction Condition Merit function: φ νk (x) = f (x) +ν k c(x) Linear model: m νk (x k, x) = f (x k ) + f (x k ) T x +ν k c(x k ) + c(x k ) T x Predicted reduction: m νk (x k ; x) = m νk (x k, 0) m νk (x k, x) Can show Dφ νk (x k ; x) m νk (x k ; x) Require: m νk (x k ; x k ) is sufficiently positive, (= guarantee direction of descent for currentν k ) or c(x k ) c(x k ) c(x k ) T x k is sufficiently positive (= controlled update ofν k to get Dφ νk (x k ; x k )<0) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

65 Theoretical Convergence Result Assumption f (x), c(x), f (x), c(x) are bounded and Lipschitz continuous, (unmodified) W k are bounded. Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

66 Theoretical Convergence Result Assumption f (x), c(x), f (x), c(x) are bounded and Lipschitz continuous, (unmodified) W k are bounded. Theorem (Curtis, Nocedal, W, 2009) If all limit points of c(x k ) have full rank thenν k is bounded and ( ) f (xk ) + c(x lim k )y k k c(x k ) = 0. Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

67 Theoretical Convergence Result Assumption f (x), c(x), f (x), c(x) are bounded and Lipschitz continuous, (unmodified) W k are bounded. Theorem (Curtis, Nocedal, W, 2009) If all limit points of c(x k ) have full rank thenν k is bounded and ( ) f (xk ) + c(x lim k )y k k c(x k ) = 0. Otherwise lim k c(x k )c(x k ) = 0 (stationary for: min 1 2 c(x) 2 2) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

68 Theoretical Convergence Result Assumption f (x), c(x), f (x), c(x) are bounded and Lipschitz continuous, (unmodified) W k are bounded. Theorem (Curtis, Nocedal, W, 2009) If all limit points of c(x k ) have full rank thenν k is bounded and ( ) f (xk ) + c(x lim k )y k k c(x k ) = 0. Otherwise lim k c(x k )c(x k ) = 0 (stationary for: min 1 2 c(x) 2 2) and ifν k is bounded lim k f (x k ) + c(x k )y k = 0. Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

69 Theoretical Convergence Result Assumption f (x), c(x), f (x), c(x) are bounded and Lipschitz continuous, (unmodified) W k are bounded. Theorem (Curtis, Nocedal, W, 2009) If all limit points of c(x k ) have full rank thenν k is bounded and ( ) f (xk ) + c(x lim k )y k k c(x k ) = 0. Otherwise lim k c(x k )c(x k ) = 0 (stationary for: min 1 2 c(x) 2 2) and ifν k is bounded lim k f (x k ) + c(x k )y k = 0. Similar result for interior point method with inequalities (Curtis, Schenk, W, 2010) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

70 Inequality Constraints / Barrier Problem p min f (x) µ ln(s (i) ) x R n,s R p i=1 s.t. c(x) = 0 d(x) + s = 0 (s> 0) W k 0 c(x k ) d(x k ) x k f (x k ) + c(x k )y 0 µs 2 k c k 0 I s + d(x k)y d k k c(x k ) T y c d(x k ) T k = µs 1 k e c(x k ) I 0 0 yk d d(x k ) + s k Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

71 Inequality Constraints / Barrier Problem p min f (x) µ ln(s (i) ) x R n,s R p i=1 s.t. c(x) = 0 d(x) + s = 0 (s> 0) W k 0 c(x k ) d(x k ) x k f (x k ) + c(x k )y 0 µi 0 S k S 1 k c + d(x k)y d k c(x k ) T k s k y c d(x k ) T k = µe c(x k ) S k 0 0 yk d d(x k ) + s k Trick: Scale slack variables: s = Sk 1 s k Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

72 Inequality Constraints / Barrier Problem p min f (x) µ ln(s (i) ) x R n,s R p i=1 s.t. c(x) = 0 d(x) + s = 0 (s> 0) W k 0 c(x k ) d(x k ) x k f (x k ) + c(x k )y 0 µi 0 S k S 1 k c + d(x k)y d k c(x k ) T k s k y c d(x k ) T k = µe c(x k ) S k 0 0 yk d d(x k ) + s k Trick: Scale slack variables: s = Sk 1 s k Now all quantities in linear system are uniformly bounded Much of equality-only analysis applies, use same termination tests Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

73 Inequality Constraints / Barrier Problem p min f (x) µ ln(s (i) ) x R n,s R p i=1 s.t. c(x) = 0 d(x) + s = 0 (s> 0) W k 0 c(x k ) d(x k ) x k f (x k ) + c(x k )y 0 µi 0 S k S 1 k c + d(x k)y d k c(x k ) T k s k y c d(x k ) T k = µe c(x k ) S k 0 0 yk d d(x k ) + s k Trick: Scale slack variables: s = Sk 1 s k Now all quantities in linear system are uniformly bounded Much of equality-only analysis applies, use same termination tests Special consideration Fraction to boundary: sk +α k s k (1 τ)s k > 0 Slack reset: sk+1 max{s k+1, d(x k+1 )} Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

74 Some Computational Results Ipopt [W, Laird, Biegler] primal-dual interior point method originally: filter line-search method added inexact algorithm with penalty function Pardiso [Schenk, Gärtner] Parallel Direct Linear Solver includes iterative linear solvers (e.g., SQMR) preconditioner: multi-level incomplete factorization with weighted graph matchings and inverse-based pivoting Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

75 Some Computational Results Ipopt [W, Laird, Biegler] primal-dual interior point method originally: filter line-search method added inexact algorithm with penalty function Pardiso [Schenk, Gärtner] Parallel Direct Linear Solver includes iterative linear solvers (e.g., SQMR) preconditioner: multi-level incomplete factorization with weighted graph matchings and inverse-based pivoting Experiments: Robustness test with CUTE/COPS collection 684 AMPL models Inexact method solved 85% (original Ipopt 94%) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

76 Some Computational Results Ipopt [W, Laird, Biegler] primal-dual interior point method originally: filter line-search method added inexact algorithm with penalty function Pardiso [Schenk, Gärtner] Parallel Direct Linear Solver includes iterative linear solvers (e.g., SQMR) preconditioner: multi-level incomplete factorization with weighted graph matchings and inverse-based pivoting Experiments: Robustness test with CUTE/COPS collection 684 AMPL models Inexact method solved 85% (original Ipopt 94%) PDE-constrained optimization problems Implemented in Matlab; finite differences; uniform grids Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

77 Boundary Control 1 (y(x) y t (x)) 2 dx 2 Ω s.t. (e y(x) y(x)) = 20 in Ω = [0, 1] 3 min y,u y(x) = u(x) on Ω 2.5 u(x) 3.5 on Ω Target profile: y t (x) = x (1) (x (1) 1)x (2) (x (2) 1) sin(2πx (3) ) N # var # eq # ineq # iter CPU sec (per iter) (0.2165) (7.9731) (380.88) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

78 Boundary Control 1 (y(x) y t (x)) 2 dx 2 Ω s.t. (e y(x) y(x)) = 20 in Ω = [0, 1] 3 min y,u y(x) = u(x) on Ω 2.5 u(x) 3.5 on Ω Target profile: y t (x) = x (1) (x (1) 1)x (2) (x (2) 1) sin(2πx (3) ) N # var # eq # ineq # iter CPU sec (per iter) (0.2165) (7.9731) (380.88) Original Ipopt with N = 32 requires 238 seconds per iteration Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

79 Hyperthermia Treatment Planning 1 (y(x) y t (x)) 2 dx 2 Ω s.t. y(x) 10(y(x) 37) = u M(x)u in Ω = [0, 1] 3 min y,u 37.0 y(x) 37.5 on Ω 42.0 y(x) 44.0 in Ω Tumor Ω Controls: u j = a j e iφ j j = 1,...,10 Data: M jk (x) =< E j (x), E k (x)>, E j = sin(jπx 1 x 2 x 3 ) N # var # eq # ineq # iter CPU sec (per iter) (0.3367) (59.920) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

80 Hyperthermia Treatment Planning 1 (y(x) y t (x)) 2 dx 2 Ω s.t. y(x) 10(y(x) 37) = u M(x)u in Ω = [0, 1] 3 min y,u 37.0 y(x) 37.5 on Ω 42.0 y(x) 44.0 in Ω Tumor Ω Controls: u j = a j e iφ j j = 1,...,10 Data: M jk (x) =< E j (x), E k (x)>, E j = sin(jπx 1 x 2 x 3 ) N # var # eq # ineq # iter CPU sec (per iter) (0.3367) (59.920) Original Ipopt with N = 32 requires 408 seconds per iteration Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

81 Inverse Problem min y,u 1 6 (y i (x) y i,t (x)) 2 dx + 1 (β(u(x) u t (x)) 2 + (u(x) u t (x)) 2) 2 2 i=1 Ω Ω s.t. (e u(x) y i (x)) = q i (x) in Ω = [0, 1] 3, i = 1,...,6 y i (x) n = 0 on Ω, i = 1,...,6 y i (x) dx = 0 i = 1,...,6 Ω 1 u(x) 1 in Ω N # var # eq # ineq # iter CPU sec (per iter) ( ) ( ) ( ) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

82 Inverse Problem min y,u 1 6 (y i (x) y i,t (x)) 2 dx + 1 (β(u(x) u t (x)) 2 + (u(x) u t (x)) 2) 2 2 i=1 Ω Ω s.t. (e u(x) y i (x)) = q i (x) in Ω = [0, 1] 3, i = 1,...,6 y i (x) n = 0 on Ω, i = 1,...,6 y i (x) dx = 0 i = 1,...,6 Ω 1 u(x) 1 in Ω N # var # eq # ineq # iter CPU sec (per iter) ( ) ( ) ( ) Original Ipopt with N = 32 requires 15 hours for the first iteration (reduction by factor 550) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

83 PDE-Constrained Optimization Very active research area Often try to use of preconditioners for original PDE Extremely efficient in solving linear systems for state variables Optimization methods often based on step decomposition Projections into the null space of constraint Jacobian Work with reduced Hessian Those often lead to triple iterative loops Can handle constraints on control variables Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

84 PDE-Constrained Optimization Very active research area Often try to use of preconditioners for original PDE Extremely efficient in solving linear systems for state variables Optimization methods often based on step decomposition Projections into the null space of constraint Jacobian Work with reduced Hessian Those often lead to triple iterative loops Can handle constraints on control variables Potential difficulty: Inequalities involving state variables If interior-point method is used: Reduced Hessian becomes very ill-conditioned Not clear how to find efficient preconditioner Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

85 Our Ongoing Work We hope that our approach can make a difference for problems with many state constraints when no efficient PDE preconditioner is known (e.g., some version of Helmholtz equation) Find an interesting problem with many state constraints Utilize standard PDE techniques Discretization using meshes (e.g., finite elements) Adaptive mesh refinement (e.g., [Ziems, Ulbrich]) Scalable distributed-memory parallel implementation If successful, distribute code Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

86 Example: Server Room Cooling Problem: Heat generating equipment in room must be cooled Want to place/control air conditioners to minimize cooling costs Suggested by Henderson, Lopez, Mello (IBM Research) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

87 Formulation Of Airflow Problem We assume Air flow has no internal friction (irrotational) Air speed is far below speed of sound (incompressible) Then air flow can be described by velocity potential Φ Air velocity: v(x) = Φ(x) Potential satisfies Laplace s equation Φ(x) = n i=1 2 Φ xi 2 (x) = 0 for x Ω Ω R n : interior of server room without equipment (n = 2, 3) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

88 Boundary conditions Φ(x) = 0 for x Ω Walls and equipment: No airflow through the surface Φ(x) n = 0 for x Γ wall ( n : outward normal derivative) At AC locations in walls: Controllable fan speed is incoming flow Φ(x) n = v ACi for x Γ ACi At exhausts: Assume hole in the wall ( grounded potential ) Φ(x) = 0 for x Γ Exhj Unique solution Φ H 1 (Ω), continuously dependent on v ACi Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

89 Optimal Control Problem min ci v ACi s.t. Φ(x) = 0 for x Ω Φ(x) = 0 n for x Γ wall Φ(x) = v ACi n for x Γ ACi Φ(x) = 0 for x Γ Exhj Φ(x) 2 2 vmin 2 v ACi 0 for x Γ hot Objective: Minimize costs of running ACs Constraints: Require minimum air speed at heat sources of equipment (nonconvex) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

90 Software Finite Element Library: libmesh [Kirk, Peterson, Stogner, Carey] Manages the finite-element mesh Interfaces to mesh generators (we use tetgen [Si, Gärtner] for 3D) Handles refinement of mesh (not yet used) Distributed memory implementation (uses PETSc) Optimization code: Ipopt (inexact algorithm) Variation: Step decomposition only when necessary Using the flexible penalty function [Curtis, Nocedal, 2008] Using distributed memory version of Ipopt [Dash, W] handles linear algebra objects in parallel supports parallel function evaluations not (yet... ) officially released Linear Solver: Pardiso (iterative linear solver) No distributed memory implementation In the future: use PSPIKE [Samel, Sathe, Schenk] Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

91 3D Server Room Cooling Example 7 servers of different sizes 57 ACs located all around the walls One exhaust in top front Constraint: Minimum air speed at front and back sides of servers Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

92 Computational Results Problem size (Ω = [0, 10] [0, 5] [0, 3]) h #var #eq #ineq #nnz KKT ,816 43,759 4, , , ,134 19,128 5,759,809 Exact algorithm h #iter f (x ) CPU(algo) CPU(eval) fill-in , , , Inexact algorithm h #iter f (x ) CPU(algo) CPU(eval) , , , (different ACs active for h = 0.1 and h = 0.05) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

93 Optimal Solution (h = 0.05) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

94 Optimal Solution (h = 0.05) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

95 Optimal Solution (h = 0.05) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

96 Optimal Solution (Active Constraints) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

97 Future Plans These computations are still preliminary Further improvements: Include temperature into the model Nonlinear PDEs Include adaptive mesh refinement within the optimization Improve preconditioner Perform computations in parallel Function evaluations (libmesh) and Ipopt are already parallel Explore distributed memory iterative linear solver PSPIKE Potentially release code as part of Ipopt distribution Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41

98 Conclusions Nonlinear optimization with inexact step computations Termination tests for iterative linear solver Consider optimization nature Based on model of merit function Theoretical convergence properties under mild assumptions Encouraging results for PDE constrained problems Very good speed up on some examples Promising preliminary results for problem with state constraints A lot left to do (Looking for other NLP applications with large fill-in ideas?) Andreas Wächter (IBM) NLP With Inexact Steps IPAM / 41