NUMERICAL OPTIMIZATION. J. Ch. Gilbert

Similar documents
MS&E 318 (CME 338) Large-Scale Numerical Optimization

How the augmented Lagrangian algorithm can deal with an infeasible convex quadratic optimization problem. Motivation, analysis, implementation

SF2822 Applied Nonlinear Optimization. Preparatory question. Lecture 9: Sequential quadratic programming. Anders Forsgren

Algorithms for Constrained Optimization

Motivation, analysis, implementation

An Inexact Sequential Quadratic Optimization Method for Nonlinear Optimization


Numerical Optimization Theoretical and Practical Aspects

Infeasibility Detection and an Inexact Active-Set Method for Large-Scale Nonlinear Optimization

Multidisciplinary System Design Optimization (MSDO)

Lecture 13: Constrained optimization

What s New in Active-Set Methods for Nonlinear Optimization?

Recent Adaptive Methods for Nonlinear Optimization

AM 205: lecture 19. Last time: Conditions for optimality, Newton s method for optimization Today: survey of optimization methods

A Trust-region-based Sequential Quadratic Programming Algorithm

Perturbation Analysis of Optimization Problems

Algorithms for constrained local optimization

5 Handling Constraints

An Inexact Newton Method for Optimization

AM 205: lecture 19. Last time: Conditions for optimality Today: Newton s method for optimization, survey of optimization methods

REGULARIZED SEQUENTIAL QUADRATIC PROGRAMMING METHODS

Infeasibility Detection in Nonlinear Optimization

arxiv: v1 [math.na] 8 Jun 2018

A New Penalty-SQP Method

Introduction. New Nonsmooth Trust Region Method for Unconstraint Locally Lipschitz Optimization Problems

1 Computing with constraints

Journal of Convex Analysis Vol. 14, No. 2, March 2007 AN EXPLICIT DESCENT METHOD FOR BILEVEL CONVEX OPTIMIZATION. Mikhail Solodov. September 12, 2005

The Squared Slacks Transformation in Nonlinear Programming

E5295/5B5749 Convex optimization with engineering applications. Lecture 8. Smooth convex unconstrained and equality-constrained minimization

Optimization Problems with Constraints - introduction to theory, numerical Methods and applications

4TE3/6TE3. Algorithms for. Continuous Optimization

A Trust Funnel Algorithm for Nonconvex Equality Constrained Optimization with O(ɛ 3/2 ) Complexity

NONLINEAR. (Hillier & Lieberman Introduction to Operations Research, 8 th edition)

Pacific Journal of Optimization (Vol. 2, No. 3, September 2006) ABSTRACT

Solving MPECs Implicit Programming and NLP Methods

MS&E 318 (CME 338) Large-Scale Numerical Optimization

Survey of NLP Algorithms. L. T. Biegler Chemical Engineering Department Carnegie Mellon University Pittsburgh, PA

Image restoration. An example in astronomy

10 Numerical methods for constrained problems

An Inexact Newton Method for Nonlinear Constrained Optimization

Nonlinear Optimization Solvers

Optimality, Duality, Complementarity for Constrained Optimization

Lecture 11 and 12: Penalty methods and augmented Lagrangian methods for nonlinear programming

PDE-Constrained and Nonsmooth Optimization

5.6 Penalty method and augmented Lagrangian method

The use of second-order information in structural topology optimization. Susana Rojas Labanda, PhD student Mathias Stolpe, Senior researcher

On the complexity of an Inexact Restoration method for constrained optimization

Primal-dual relationship between Levenberg-Marquardt and central trajectories for linearly constrained convex optimization

Part 5: Penalty and augmented Lagrangian methods for equality constrained optimization. Nick Gould (RAL)

Trust-Region SQP Methods with Inexact Linear System Solves for Large-Scale Optimization

Penalty and Barrier Methods. So we again build on our unconstrained algorithms, but in a different way.

Numerical Optimization Professor Horst Cerjak, Horst Bischof, Thomas Pock Mat Vis-Gra SS09

Sequential Quadratic Programming Method for Nonlinear Second-Order Cone Programming Problems. Hirokazu KATO

Lecture 15: SQP methods for equality constrained optimization

5.5 Quadratic programming

COMPARATIVE STUDY BETWEEN LEMKE S METHOD AND THE INTERIOR POINT METHOD FOR THE MONOTONE LINEAR COMPLEMENTARY PROBLEM

Advanced Continuous Optimization

Optimization. Escuela de Ingeniería Informática de Oviedo. (Dpto. de Matemáticas-UniOvi) Numerical Computation Optimization 1 / 30

Scientific Computing: An Introductory Survey

Scientific Computing: An Introductory Survey

Determination of Feasible Directions by Successive Quadratic Programming and Zoutendijk Algorithms: A Comparative Study

Nonlinear Programming, Elastic Mode, SQP, MPEC, MPCC, complementarity

AN EXACT PENALTY APPROACH FOR MATHEMATICAL PROGRAMS WITH EQUILIBRIUM CONSTRAINTS. L. Abdallah 1 and M. Haddou 2

Preprint ANL/MCS-P , Dec 2002 (Revised Nov 2003, Mar 2004) Mathematics and Computer Science Division Argonne National Laboratory

Nonlinear Optimization: What s important?

LARGE SCALE NONLINEAR OPTIMIZATION

A Regularized Interior-Point Method for Constrained Nonlinear Least Squares

An Active Set Strategy for Solving Optimization Problems with up to 200,000,000 Nonlinear Constraints

CE 191: Civil and Environmental Engineering Systems Analysis. LEC 05 : Optimality Conditions

Inexact Newton Methods and Nonlinear Constrained Optimization

Optimization Concepts and Applications in Engineering

Optimization. Yuh-Jye Lee. March 28, Data Science and Machine Intelligence Lab National Chiao Tung University 1 / 40

Derivative-Based Numerical Method for Penalty-Barrier Nonlinear Programming

Exam 2. Average: 85.6 Median: 87.0 Maximum: Minimum: 55.0 Standard Deviation: Numerical Methods Fall 2011 Lecture 20

Part 4: Active-set methods for linearly constrained optimization. Nick Gould (RAL)

Outline. Scientific Computing: An Introductory Survey. Optimization. Optimization Problems. Examples: Optimization Problems

On the Local Quadratic Convergence of the Primal-Dual Augmented Lagrangian Method

CONVERGENCE ANALYSIS OF AN INTERIOR-POINT METHOD FOR NONCONVEX NONLINEAR PROGRAMMING

Computational Optimization. Augmented Lagrangian NW 17.3

Numerical Methods for PDE-Constrained Optimization

Rachid Benouahboun 1 and Abdelatif Mansouri 1

A Piecewise Line-Search Technique for Maintaining the Positive Definiteness of the Matrices in the SQP Method

2.3 Linear Programming

Seminal papers in nonlinear optimization

Advanced Continuous Optimization

Numerical Nonlinear Optimization with WORHP

Constrained Nonlinear Optimization Algorithms

Iterative Reweighted Minimization Methods for l p Regularized Unconstrained Nonlinear Programming

Solution Methods. Richard Lusby. Department of Management Engineering Technical University of Denmark

8 Numerical methods for unconstrained problems

LINEAR AND NONLINEAR PROGRAMMING

A STABILIZED SQP METHOD: SUPERLINEAR CONVERGENCE

Numerisches Rechnen. (für Informatiker) M. Grepl P. Esser & G. Welper & L. Zhang. Institut für Geometrie und Praktische Mathematik RWTH Aachen

Introduction to Nonlinear Stochastic Programming

Interior Point Methods for Convex Quadratic and Convex Nonlinear Programming

A GLOBALLY CONVERGENT STABILIZED SQP METHOD: SUPERLINEAR CONVERGENCE

Motivation. Lecture 2 Topics from Optimization and Duality. network utility maximization (NUM) problem:

A New Low Rank Quasi-Newton Update Scheme for Nonlinear Programming

A semidefinite relaxation scheme for quadratically constrained quadratic problems with an additional linear constraint

Lecture 3. Optimization Problems and Iterative Algorithms

Transcription:

NUMERICAL OPTIMIZATION J. Ch. Gilbert

Numerical optimization (past) The discipline deals with the classical smooth (nonconvex) problem min {f(x) : c E (x) = 0, c I (x) 0}. Applications: variable added lens design (Essilor), seismic tomography (IFP), tire industry (Michelin). Contributions to interior point (IP) methods Technique: solving min {f(x) µ P m i=1 log s i : c E (x) = 0, c I (x) + s = 0 for µ 0. Theoretical convergence studies of the quasi-newton version of the approach (with P. Armand, S. Jégou). Introduction and theoretical convergence study of one of the 1st nonlinear IP approach (with R. Byrd, J. Nocedal). Ill-behaved central paths for convex problems (with E. Karas, C. Gonzaga). Development of the prorotype general purpose solver OPINL (with A. Fuduli). Contributions to the SQP (= Newton) approach Technique: solving a sequence of (convex) QP s min {g d+ 1 2 d Hd : c E (x)+c E (x)d = 0, c I(x)+c I (x)d 0}. Truncated-Newton version with line-search (with L. Chauvier, A. Fuduli). Global linear convergence of the AL approach for solving a convex QP (with F. Delbos). Development of the general SQP solver SQPAL using the AL for solving QP s (in / with IFP). PhD thesis: E. Karas (2002), X. Jonsson (2002), F. Delbos (2004). A book: Numerical Optimization: Theoretical and Practical Aspects, J.F. Bonnans, J.Ch. Gilbert, C. Lemaréchal, and C.A. Sagastizábal, Springer, 2003. J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004 2

Numerical optimization: back to SQP The discipline deals with the classical smooth (nonconvex) problem (P EI ) < : min x f(x) c E (x) = 0 c I (x) 0. Applications: variable added lens design (Essilor), seismic tomography (IFP), tire industry (Michelin), potentially many more... The sequential quadratic programming (SQP) approach (mid 70) It is a Newton-like method. It solves a sequence of QP s. < : min d g d + 1 2 d Hd c E (x) + c E (x)d = 0 c I (x) + c I (x)d 0. Properties: few iterations, QP expensive to solve. Interior point (IP) approach (late 90) It is a penalty method. It solves a sequence of nonlinear OP s without inequalities. < : min x f(x) µ P m i=1 log s i c E (x) = 0 c I (x) + s = 0. Properties: more iterations, LS less expensive to solve, ill-conditioned LS. For very large problems: SQP looks better that IP! J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004 3

Solving a quadratic optimization problem (QP) The problem to solve Let Q S n and C be m n. (QP) j minx 1 2 x Qx + q x l Cx u. Difficulties: inequality constraints Q 0 = NP-hardness Standard approach within SQP (e.g., SNOPT) active set method Solves for a sequence of updated working set W {1,..., m} and implicit bounds l Cx u: j minx 1 2 x Qx + q x C i x = l i or u i, for i W. Properties: finite termination, W changes slowly. Forgotten approach augmented Lagrangian (AL) method (requires Q 0) Solves for a sequence of multipliers updated by λ + = λ + r(cx + y + ): (ALP) j 1 minx 2 x Qx + q x + λ (Cx y) + r 2 Cx y 2 l y u. Properties: (ALP) solved by GP-AS-CG (for example): efficient since only bound constraints and FIAC property, r can change at each iteration, infinite termination, but fast convergence with a global convergence rate: for some L > 0, at each iteration Cx + y + Cx y min 1, L «. r J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004 4

Solving large scale seismic tomography problems with SQPAL QPAL and SQPAL QPAL: solver of (QP) with the AL approach (Fortran-90 version at IFP, Matlab version at Inria). SQPAL: SQP algorithm to solve (P EI ) using QPAL as QP solver. A seismic tomography problem (with F. Delbos, R. Glowinski, D. Sinoquet) Can solve problems with up to n 1.5 10 4 variables and up to m 10 4 constraints. J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004 5

Numerical experiments with QPAL ncg 10 5 10 4 Experiments for n = 1,..., 10 3, m = max(1, n/2 ), and random data. 10 3 10 2 10 1 Total number of CG iterations n CG in terms of n. the behavior looks linear : n CG 6.67n 1.16 (n CG O(n 1.5 ) for IP). 10 0 10 0 10 1 10 2 10 3 n 12 ncgrel 10 6 4 The figure gives n CG w.r.t. n CG for the unconstrained problem. The ratio looks bounded (O(n 0.5 ) for IP). 2 0 10 0 10 1 10 2 10 3 n J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004 6

The future of QPAL Open questions Behavior when the QP is infeasible (extension of a work by R. Glowinski). Solving inexactly the bound constrained augmented Lagrangian (see Dostál, Friedlander, and Santos). Make more precise the finite identification property of the AL approach. Use of QPAL in SQP with trust regions. Polynomial complexity of the algorithm? (suggested by the previous numerical experiments) Comparison with IP. Can we say something when the QP is nonconvex? (... close to P = NP!) Possible application to shape optimization (aeronautic and tire industry) and many other domains. J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004 7

Optimization problems with complementarity constraints (MPEC) MPEC (mathematical programming with equilibrium constraints): an intensive international activity on the problem < : min x f(x) c I (x) = 0 and c I (x) 0 0 p(x) q(x) 0 complementarity constraints. Difficulty: complementarity constraints are not qualified! Numerical methods IP approach: already widely explored. Renewal of the SQP approach (for some classes of problems). Does a sequential QP+(linear complementarity constraint) can be used? Can an augmented Lagrangian approach be useful? Possible applications Robotics, contact problems (tire industry, Michelin),... Optimization of many systems with threshold effects: stratigraphic identification (IFP),... Special cases of bilevel optimization problems: yield management (in networks, France Télécom),... J.Ch. Gilbert INRIA-Rocquencourt October 14, 2004