( x) min. Nonlinear optimization problem without constraints NPP: then. Global minimum of the function f(x)

Similar documents
7.0 Equality Contraints: Lagrange Multipliers

LINEARLY CONSTRAINED MINIMIZATION BY USING NEWTON S METHOD

A Penalty Function Algorithm with Objective Parameters and Constraint Penalty Parameter for Multi-Objective Programming

A New Method for Solving Fuzzy Linear. Programming by Solving Linear Programming

Some Different Perspectives on Linear Least Squares

An Alternative Approach to the Solution of Multi-Objective Geometric Programming Problems

Capacitated Plant Location Problem:

A conic cutting surface method for linear-quadraticsemidefinite

-Pareto Optimality for Nondifferentiable Multiobjective Programming via Penalty Function

C.11 Bang-bang Control

= 2. Statistic - function that doesn't depend on any of the known parameters; examples:

Algorithms behind the Correlation Setting Window

M2S1 - EXERCISES 8: SOLUTIONS

D. L. Bricker, 2002 Dept of Mechanical & Industrial Engineering The University of Iowa. CPL/XD 12/10/2003 page 1

Generalized Linear Regression with Regularization

MULTIOBJECTIVE NONLINEAR FRACTIONAL PROGRAMMING PROBLEMS INVOLVING GENERALIZED d - TYPE-I n -SET FUNCTIONS

1 Lyapunov Stability Theory

Binary classification: Support Vector Machines

UNIT 7 RANK CORRELATION

A nonsmooth Levenberg-Marquardt method for generalized complementarity problem

Support vector machines

The Mathematical Appendix

( ) ( ) ( ( )) ( ) ( ) ( ) ( ) ( ) = ( ) ( ) + ( ) ( ) = ( ( )) ( ) + ( ( )) ( ) Review. Second Derivatives for f : y R. Let A be an m n matrix.

Support vector machines II

Solutions to Odd-Numbered End-of-Chapter Exercises: Chapter 17

CS 2750 Machine Learning Lecture 8. Linear regression. Supervised learning. a set of n examples

An Implementation of Integer Programming Techniques in Clustering Algorithm

A COMPARATIVE STUDY OF THE METHODS OF SOLVING NON-LINEAR PROGRAMMING PROBLEM

Global Optimization for Solving Linear Non-Quadratic Optimal Control Problems

CS 1675 Introduction to Machine Learning Lecture 12 Support vector machines

Review Exam I Complex Analysis. Cauchy s Integral Formula (#0). Let G be a region in C, let Bar (, ) G and let γ be the circle C(a,r), oriented.

Hamilton s principle for non-holonomic systems

Numerical Experiments with the Lagrange Multiplier and Conjugate Gradient Methods (ILMCGM)

ELEMENTS OF NUMBER THEORY. In the following we will use mainly integers and positive integers. - the set of integers - the set of positive integers

Debabrata Dey and Atanu Lahiri

KURODA S METHOD FOR CONSTRUCTING CONSISTENT INPUT-OUTPUT DATA SETS. Peter J. Wilcoxen. Impact Research Centre, University of Melbourne.

Multiple Choice Test. Chapter Adequacy of Models for Regression

Arithmetic Mean and Geometric Mean

Strong Convergence of Weighted Averaged Approximants of Asymptotically Nonexpansive Mappings in Banach Spaces without Uniform Convexity

COV. Violation of constant variance of ε i s but they are still independent. The error term (ε) is said to be heteroscedastic.

0/1 INTEGER PROGRAMMING AND SEMIDEFINTE PROGRAMMING

An Algorithm for Capacitated n-index Transportation Problem

Assignment 5/MATH 247/Winter Due: Friday, February 19 in class (!) (answers will be posted right after class)

Advances in Theory of Neural Network and Its Application

OPTIMALITY CONDITIONS FOR LOCALLY LIPSCHITZ GENERALIZED B-VEX SEMI-INFINITE PROGRAMMING

Exercises for Square-Congruence Modulo n ver 11

Functions of Random Variables

Graphs and graph models-graph terminology and special types of graphs-representing graphs and graph isomorphism -connectivity-euler and Hamilton

Order Nonlinear Vector Differential Equations

Maximum Walk Entropy Implies Walk Regularity

Special Instructions / Useful Data

Chapter 3. Differentiation 3.2 Differentiation Rules for Polynomials, Exponentials, Products and Quotients

A Conventional Approach for the Solution of the Fifth Order Boundary Value Problems Using Sixth Degree Spline Functions

Lecture 07: Poles and Zeros

Numerical Analysis Formulae Booklet

Numerical Simulations of the Complex Modied Korteweg-de Vries Equation. Thiab R. Taha. The University of Georgia. Abstract

A CLASS OF SINGULAR PERTURBATED BILOCAL LINEAR PROBLEMS

Linear regression (cont) Logistic regression

Chapter 3. Differentiation 3.3 Differentiation Rules

On the Pfaff problem

CSE 5526: Introduction to Neural Networks Linear Regression

Solutions to problem set ); (, ) (

Given a table of data poins of an unknown or complicated function f : we want to find a (simpler) function p s.t. px (

ROOT-LOCUS ANALYSIS. Lecture 11: Root Locus Plot. Consider a general feedback control system with a variable gain K. Y ( s ) ( ) K

1 Onto functions and bijections Applications to Counting

α1 α2 Simplex and Rectangle Elements Multi-index Notation of polynomials of degree Definition: The set P k will be the set of all functions:

LINEAR REGRESSION ANALYSIS

UNIT 6 CORRELATION COEFFICIENT

this is the indefinite integral Since integration is the reverse of differentiation we can check the previous by [ ]

ECE 421/599 Electric Energy Systems 7 Optimal Dispatch of Generation. Instructor: Kai Sun Fall 2014

Ideal multigrades with trigonometric coefficients

Investigation of Partially Conditional RP Model with Response Error. Ed Stanek

A MODIFIED REGULARIZED NEWTON METHOD FOR UNCONSTRAINED NONCONVEX OPTIMIZATION

A Note on the Almost Sure Central Limit Theorem in the Joint Version for the Maxima and Partial Sums of Certain Stationary Gaussian Sequences *

DUALITY FOR MINIMUM MATRIX NORM PROBLEMS

An Introduction to. Support Vector Machine

Point Estimation: definition of estimators

1 0, x? x x. 1 Root finding. 1.1 Introduction. Solve[x^2-1 0,x] {{x -1},{x 1}} Plot[x^2-1,{x,-2,2}] 3

Power Flow S + Buses with either or both Generator Load S G1 S G2 S G3 S D3 S D1 S D4 S D5. S Dk. Injection S G1

5. Data Compression. Review of Last Lecture. Outline of the Lecture. Course Overview. Basics of Information Theory: Markku Juntti

Neural Networks for Nonlinear Fractional Programming

1 Solution to Problem 6.40

CHAPTER 4 RADICAL EXPRESSIONS

Regression and the LMS Algorithm

Unsupervised Learning and Other Neural Networks

Cesaro Supermodular Order and Archimedean Copulas

Dr. Shalabh. Indian Institute of Technology Kanpur

Non-degenerate Perturbation Theory

3D Reconstruction from Image Pairs. Reconstruction from Multiple Views. Computing Scene Point from Two Matching Image Points

Bayes (Naïve or not) Classifiers: Generative Approach

Correlation and Regression Analysis

Chapter 9 Jordan Block Matrices

Relations to Other Statistical Methods Statistical Data Analysis with Positive Definite Kernels

ALGORITHMS FOR OPTIMAL DECISIONS 2. PRIMAL-DUAL INTERIOR POINT LP ALGORITHMS 3. QUADRATIC PROGRAMMING ALGORITHMS & PORTFOLIO OPTIMISATION

INVARIANT MANIFOLDS OF COMPLEX SYSTEMS

The Necessarily Efficient Point Method for Interval Molp Problems

Applied Fitting Theory VII. Building Virtual Particles

Lyapunov Stability. Aleksandr Mikhailovich Lyapunov [1] 1 Autonomous Systems. R into Nonlinear Systems in Mechanical Engineering Lesson 5

Summary of the lecture in Biostatistics

are positive, and the pair A, B is controllable. The uncertainty in is introduced to model control failures.

Transcription:

Objectve fucto f() : he optzato proble cossts of fg a vector of ecso varables belogg to the feasble set of solutos R such that It s eote as: Nolear optzato proble wthout costrats NPP: R f ( ) : R R f f R f ( ) ( ) = f Local u a global u of the fucto f() he vector s local u of the fucto f() the space R f there ests such ope space E R of a pot that E f ( ) f ( ) Atoaly f f ( ) < f ( ) for the there ests strct local u. he vector s global u of the fucto f() the space R f there ests such ope space R of a pot that R f ( ) f ( ) Atoaly f f ( ) < f ( ) for the there ests the strct global u ths pot. Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Nolear optzato proble wthout costrats: Defto. A recto the space R we eote ay -etoal colu vector. Let us assue that there ests a pot R a scalar τ [ ; + ). Ay pot y R Optzato theory a coputg belogg to half strght le whch starts pot recto wll be etere as follows: Lea. Let f : X= R R y= + τ be a fferetable fucto the pot Let us assue that there est a recto for whch: ( ) < X he there ests such σ > that for all τ ( σ ] the followg coto s fulflle f ( +τ) < f ( ). Proof: let us use the property of Gateau fferetal wth the rectoal ervatve efto. Faculty of Electrocs Nolear optzato proble wthout costrats: heore. Let f : X= R R wll be fferetable fucto. If ze fucto f() the for: f ( ) f ( ) X Proof: ot rectly. he pot f ( ˆ) f ( ) For each X f a oly f whe fulflls the followg coto: Optzato theory a coputg the (ˆ) = s ae as a statoary pot. heore. Let f : X = R R wll be cove a fferetable fucto. For the pot ˆ X fucto f() costtutes al value of the fucto: ( ˆ) = It s the ecessary coto for local etreu f() the pot. X Faculty of Electrocs Global u of the fucto f() Suffcat cotos for the olear optzato proble wthout costrats: Fucto f() s cotuous a twce fferetable. he fucto f() has the atr wth seco ervatves ae as hesja A heore: Matr A has soe a subeterats A If the fucto f : X = R R wll be strctly cove a fferetable the the vector ˆ X whch fulflls the ecessary coto ( ˆ) = s the oly oe global u of the fucto f(). A = ( ) ( ) A =.. ( ).... ( ) ( ) ( ) A = ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs

he statoary cotos for olear prograg proble wthout costrats Let us assue that s a statoary pot for a fuctof(). he the followg relatos are fulflle:. If the hesja atr A s postve efe : A( ) > for = the a fucto f() has local u ths pot.. If the hesja atr A s egatve efe : ( ) A( ) > for = the a fucto f() has local au ths pot. 3. If the hesja s half-postve efe: or the hesja s half-egatve efe: A( ) for = a A( ) = ( ) A( ) for = a A( ) = heore. If the fucto f() s twce fferetable the each local u for the olear prograg proble wthout costrats the followg optalty cotos are fulflle: = A ( ) > A= f ( ) for I orer coto II orer coto Matr A has to be strctly postve I ths oet t s possble to ece what type of a etreu of a fucto f( s ths pot. 4. If two cotos a are ot fulflle (the the hesja A s ot etere) the the fucto f() has ot etreu the pot. Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Nolear prograg proble wth costrats NPP : Eaple of NP proble Let us f where: such that: f ( ˆ) = f ( ) X = { : g ( ) = } f : X = R g : X = R X R R. f ( ) = ( X 3 ) + 4 ( ) + ( ) X = {( ) ( ) (.4+ ) 3 3/ + ( 3 3 = )}. Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Nolear prograg proble wth costrats NPP : Defto. he recto startg the pot s feasble f there ests such σ > that for ay τ [ ; σ ] + τ X he set of all feasble rectos: D( ) = { : σ > such that τ [; σ ] + τ X }. Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs

Lagrage fucto Actve costrats: Defto. he Lagrage fucto for olear prograg proble s a scalar fucto as follows:. g (ˆ + τ) A(ˆ) where τ [ σ ] L ( λ) = f ( ) + λ g( ). g (ˆ + τ) = g (ˆ) + τ g (ˆ) + O( τ ) where λ R s a vector of Lagrage ultplers. 3. he ecessary coto for the feasblty of recto vector s: g ( ˆ) A(ˆ) Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Farkas lea he set of rectos: Let us efe the set of -esoal vectors R space { b a = }. he relato set up for each R b whch fulflls a b+ = λ a =. λ = [ λ ] λ D ( ) = : <g( ) > A( ) < ( ) D ( ) = : <g( ) > A( ) < ( ) > >< D3( ) = : <g ( ) >> forsoe A( ) ( 3 D ) = D ( ) D ( ) D ( ) Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Kuh-ucker-Karusch ecessary cotos: Proof: ecessary cotos Kuh-ucker-Karusch heore: If a)fuctos ˆ Optzato theory a coputg f g are fferetable; b) s a local u of NPP there ests λ ˆ ˆ λ =. such that f a oly f (ˆ) + g (ˆ) = (ˆ) = = = D ( ˆ) =φ Faculty of Electrocs Proof. Let Optzato theory a coputg a a =g ( ˆ) A(ˆ) he accorg to the Farkas lea there ests Such that whe A the coto for the actve costrats has to be fulflle : g ˆ) f a oly f b= (ˆ) (ˆ) = ˆ λ ( g (ˆ)) A( ˆ ) (ˆ ) R ( A(ˆ) D ( ˆ) =φ ˆ λ A(ˆ) Faculty of Electrocs

Proof: ecessary cotos Kuh-ucker-Karusch Kuh-ucker_Karush ecessary cotos wth the help of Lagraga fucto For A(ˆ) let us assue ˆ = So two followg equatos are fulflle: (ˆ) + = λ g (ˆ) = (ˆ) = = Ck L ( λ ) = f ( ) + λ Necessary cotos : L = g ( ˆ ) = λ ˆ L(ˆ ) = λ λ L ( ˆ ) ˆ λ ( ). Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Suffcet cotos efe as the regularty coto:. All fuctos for the cotrats g () are lear the Karl regularty coto... All fuctos for the cotrats g () are cove a the feasble set of solutos s ot epty - the Slater regularty coto. 3. he graets of all actve costrats g ( ) for A) are leary epeet Facco a McCorck regularty coto. ( Kuh-Karusch-ucker theore ecessary a suffcet cotos for olear prograg proble wth costrats (NPP) heore. If a)fuctos f ag are fferetable; b) s local u of NPP ˆ c) Regularty coto for costrats s fulflle the pot ˆ he there ests =. Such that the pot the Kuh-Karusch-ucker ecessary cotos are fulflle: (ˆ) + ˆ λg (ˆ) = = (ˆ) = = Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Eaple I Nolear prograg proble wth costrats NPP wth atoal costrats for varables: akg uer coserato the K-K- cotos solve the followg olear optzato proble: f ( ) = X he solutos of NPP proble: + * +.5* + X = a f (ˆ) f ( ) X = { : g ( ) = } f ( ) : R R g ( ) : R R = = X where : = [ ] f ( ) = L ( λ ) = f ( ) + λ = g ( ). Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs

L λ L λ = Lagrage cotos for NNP wth λl λ L λ = λ λ λ Eaples for Kuh-ucker-Karusch cotos Eaple II akg uer coserato the K--K cotos solve the followg olear optzato proble: f ( ) = * X + * +.5* + X = he solutos of NPP proble: = [ ] f ( ) = Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Nolear prograg proble wth costrats NPP wth atoal equatos as costrats : where : X = { : g ( ) = p h ( ) = = p+ } Where: f ( ) : X = R R g ( ) : X = R R = p h ( ) : X = R R = p+ = X f ( ) f ( ) L ( λ ) = f ( ) + λ = g ( ). heore: If a) Fuctos f a g for = are fferetable; b) s local u of NPP ˆ he there ests: λ = p a λ = p+ such that the pot are fulflle: (ˆ) + = (ˆ) = ˆ wthout ay costrats o varables the Kuh-Karusch-ucker ecessary cotos g (ˆ) = = p Optzato theory a coputg Faculty of Electrocs Optzato theory a coputg Faculty of Electrocs Eaple III akg uer coserato the K-K- cotos solve the followg olear optzato proble: f ( ) = + + X he solutos of NPP proble: 4 X = = = [ ] ; λ= f ( ) = 4 Optzato theory a coputg Faculty of Electrocs