Trust region methods: global/local convergence, approximate January methods 24, / 15

Like dokumenter
UNIVERSITETET I OSLO

SVM and Complementary Slackness

Splitting the differential Riccati equation

Graphs similar to strongly regular graphs

Level Set methods. Sandra Allaart-Bruin. Level Set methods p.1/24

UNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS

Moving Objects. We need to move our objects in 3D space.

On multigrid methods for the CahnHilliard equation with obstacle potential

Verifiable Secret-Sharing Schemes

Gradient. Masahiro Yamamoto. last update on February 29, 2012 (1) (2) (3) (4) (5)

Solutions #12 ( M. y 3 + cos(x) ) dx + ( sin(y) + z 2) dy + xdz = 3π 4. The surface M is parametrized by σ : [0, 1] [0, 2π] R 3 with.

Mathematics 114Q Integration Practice Problems SOLUTIONS. = 1 8 (x2 +5x) 8 + C. [u = x 2 +5x] = 1 11 (3 x)11 + C. [u =3 x] = 2 (7x + 9)3/2

Oppgave 1a Definer følgende begreper: Nøkkel, supernøkkel og funksjonell avhengighet.

MA2501 Numerical methods

Universitetet i Bergen Det matematisk-naturvitenskapelige fakultet Eksamen i emnet Mat131 - Differensiallikningar I Onsdag 25. mai 2016, kl.

Slope-Intercept Formula

Kneser hypergraphs. May 21th, CERMICS, Optimisation et Systèmes

melting ECMI Modelling week 2008 Modelling and simulation of ice/snow melting Sabrina Wandl - University of Linz Tuomo Mäki-Marttunen - Tampere UT

FYSMEK1110 Eksamensverksted 23. Mai :15-18:00 Oppgave 1 (maks. 45 minutt)

REMOVE CONTENTS FROM BOX. VERIFY ALL PARTS ARE PRESENT READ INSTRUCTIONS CAREFULLY BEFORE STARTING INSTALLATION

UNIVERSITETET I OSLO

Eksamen i TTK4135 Optimalisering og regulering

Generalization of age-structured models in theory and practice

On time splitting for NLS in the semiclassical regime

Neural Network. Sensors Sorter

Level-Rebuilt B-Trees

Dynamic Programming Longest Common Subsequence. Class 27

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

INF2820 Datalingvistikk V2011. Jan Tore Lønning & Stephan Oepen

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Du må håndtere disse hendelsene ved å implementere funksjonene init(), changeh(), changev() og escape(), som beskrevet nedenfor.

Existence of resistance forms in some (non self-similar) fractal spaces

Oppgave 1. ( xφ) φ x t, hvis t er substituerbar for x i φ.

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Den som gjør godt, er av Gud (Multilingual Edition)

Motzkin monoids. Micky East. York Semigroup University of York, 5 Aug, 2016

5 E Lesson: Solving Monohybrid Punnett Squares with Coding

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT


Trigonometric Substitution

TMA4329 Intro til vitensk. beregn. V2017

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

3/1/2011. I dag. Recursive descent parser. Problem for RD-parser: Top Down Space. Jan Tore Lønning & Stephan Oepen

KROPPEN LEDER STRØM. Sett en finger på hvert av kontaktpunktene på modellen. Da får du et lydsignal.

Second Order ODE's (2P) Young Won Lim 7/1/14

Hvor mye teoretisk kunnskap har du tilegnet deg på dette emnet? (1 = ingen, 5 = mye)

Bibliotekundervisningens fremtid nytt fokus på metodikk og digitalisering

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Stationary Phase Monte Carlo Methods

P(ersonal) C(omputer) Gunnar Misund. Høgskolen i Østfold. Avdeling for Informasjonsteknologi

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Hvordan føre reiseregninger i Unit4 Business World Forfatter:

UNIVERSITETET I OSLO

0:7 0:2 0:1 0:3 0:5 0:2 0:1 0:4 0:5 P = 0:56 0:28 0:16 0:38 0:39 0:23

Eksamensoppgave i TMA4135 Matematikk 4D

Norsk (English below): Guide til anbefalt måte å printe gjennom plotter (Akropolis)

TMA4245 Statistikk Eksamen 9. desember 2013

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

GYRO MED SYKKELHJUL. Forsøk å tippe og vri på hjulet. Hva kjenner du? Hvorfor oppfører hjulet seg slik, og hva er egentlig en gyro?

Eksamensoppgave i TMA4320 Introduksjon til vitenskapelige beregninger

MINIMIZATION OF LINEAR FUNCTIONALS DEFINED ON SOLUTIONS OF LARGE-SCALE DISCRETE ILL-POSED PROBLEMS

32.2. Linear Multistep Methods. Introduction. Prerequisites. Learning Outcomes

Finite Elements Methods. Formulary for Prof. Estor's exam

UNIVERSITETET I OSLO

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

ECON3120/4120 Mathematics 2, spring 2004 Problem solutions for the seminar on 5 May Old exam problems

Han Ola of Han Per: A Norwegian-American Comic Strip/En Norsk-amerikansk tegneserie (Skrifter. Serie B, LXIX)

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Ringvorlesung Biophysik 2016

(see table on right) 1,500,001 to 3,000, ,001pa to 250,000pa

EKSAMENSOPPGAVE I FAG TKP 4105

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Gir vi de resterende 2 oppgavene til én prosess vil alle sitte å vente på de to potensielt tidskrevende prosessene.

Lattice Simulations of Preheating. Gary Felder KITP February 2008

Trådløsnett med. Wireless network. MacOSX 10.5 Leopard. with MacOSX 10.5 Leopard

Windlass Control Panel

Independent Inspection

Lipschitz Metrics for Non-smooth evolutions

Ma Linær Algebra og Geometri Øving 5

Periodic solutions for a class of non-autonomous differential delay equations

Skjema for spørsmål og svar angående: Skuddbeskyttende skjold Saksnr TED: 2014/S

Du vil trenge: Umbraco Nøkkel (Inkludert); Skiftnøkkel/Fastnøkkel sett (For styrestem bolten og de 4 styreboltene),pumpe

SCE1106 Control Theory

klassisk angoragenser classic angora sweater

Perpetuum (im)mobile

INF2820 Datalingvistikk V2012. Jan Tore Lønning

HØGSKOLEN I NARVIK - SIVILINGENIØRUTDANNINGEN

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Andrew Gendreau, Olga Rosenbaum, Anthony Taylor, Kenneth Wong, Karl Dusen

Exam in Quantum Mechanics (phys201), 2010, Allowed: Calculator, standard formula book and up to 5 pages of own handwritten notes.

2003/05-001: Dynamics / Dynamikk

(see table on right) 1,500,001 to 3,000, ,001pa to 250,000pa

Se vedlagte skisse som viser hvordan det er tenkt. Dette har også grunneier bekreftet til oss vil være ok.

Emnedesign for læring: Et systemperspektiv

UNIVERSITETET I OSLO ØKONOMISK INSTITUTT

Hvor mye teoretisk kunnskap har du tilegnet deg på dette emnet? (1 = ingen, 5 = mye)

Endelig ikke-røyker for Kvinner! (Norwegian Edition)

Kartleggingsskjema / Survey

Transkript:

Trust region methods: global/local convergence, approximate methods January 24, 2014 Trust region methods: global/local convergence, approximate January methods 24, 2014 1 / 15

Trust-region idea Model m k f ( ): f (x k + p) m k (p) = f (x k ) + gk T p + 1 2 pt B k p where g k = f (x k ), B k { I, 2 f (x k ), 2 f (x k ) }. Trust-region subproblem: minimize m k (p), subject to p k. Trust region methods: global/local convergence, approximate January methods 24, 2014 2 / 15

Trust-region algorithm: 1: while g k ɛ do 2: p k := approximate solve of TR subproblem 3: Reduction: ρ k := f (x k) f (x k +p k ) m(0) m(p k ) 4: if ρ k < 1/4 then 5: k+1 = k /4 6: else if ρ k > 3/4 & p k = k then 7: k+1 = 2 k 8: else 9: k+1 = k 10: end if 11: if ρ k > η then 12: Accept step: x k+1 = x k + p k 13: else 14: Reject step: x k+1 = x k 15: end if 16: end while Trust region methods: global/local convergence, approximate January methods 24, 2014 3 / 15

Local convergence of TR-Newton algorithm Theorem 4.9: superlinear convergence of TR-Newton f C 2, 2 f -Lipschitz x k x, x satisfies 2nd order suff. cond. B k = 2 f (x k ), Newton step p N k = 2 f (x k ) 1 f (x k ) If p N k k/2 = p k p N k = o( pn k ) Then trust-region bound is inactive for all large k and x k+1 x lim k x k x = 0. Trust region methods: global/local convergence, approximate January methods 24, 2014 4 / 15

When can we expect convergence? When the subproblem is solution beats the steepest descent step (the Cauchy point p C ). Trust region methods: global/local convergence, approximate January methods 24, 2014 5 / 15

Reminder: Cauchy point calculation Cauchy point is the minimizer of m k ( αg k ), 0 α k g k. 1: if g T k B kg k 0 then 2: p C k = k g k g k 3: else 4: α = g k 2 g T k B kg k 5: if αg k k then 6: p C k = αg k 7: else 8: p C k = k g k g k 9: end if 10: end if Trust region methods: global/local convergence, approximate January methods 24, 2014 6 / 15

Global convergence of TR-algorithms Theorem 4.5, 4.6 Then: f C 2, 2 f -Lipschitz, f -bounded from below B k β, k p k k : m(p k ) m(pk C ) [can be relaxed, significantly!] 1 If any reduction is accepted by TR-alg (any ρ k > 0) = lim inf k g k = 0; 2 If only sufficient reduction is accepted (ρ k > η, η (0, 1/4)) = lim k g k = 0. Trust region methods: global/local convergence, approximate January methods 24, 2014 7 / 15

Cauchy-point based TR algorithms: Steepest descent Dogleg 2D-minimization Trust region methods: global/local convergence, approximate January methods 24, 2014 8 / 15

Steepest descent: p k = pk C Good: Cheap (explicit formula for p k ) Provably convergent Works when B k 0! Bad: SLOW! Trust region methods: global/local convergence, approximate January methods 24, 2014 9 / 15

Dogleg: Minimize m k over straght line segments: [0, g k 2 gk T B g k ] [ g k 2 kg k gk T B g k, B 1 kg k k g k]. Good: Relatively cheap (one linear system solve/tr subproblem) Provably convergent (improves upon Cauchy point) Superlinearly locally convergent (Newton steps are taken, whenever feasible) Bad: does not work when B k 0 e.g. can revert to Cauchy points in this case. Trust region methods: global/local convergence, approximate January methods 24, 2014 10 / 15

Dogleg: Path: p(τ) = p U = g 2 g T Bg g pb = B 1 g { τp U, 0 τ 1, p U + (τ 1)(p B p U ), 1 τ 2. Lemma 4.2 B k 0. Then 1 τ p(τ) monotonicaly non-decreasing; (increasing when p U 0 and p B p U ); 2 τ m( p(τ)) monotonically non-increasing. Trust region methods: global/local convergence, approximate January methods 24, 2014 11 / 15

Dogleg: 1: if B k 0 then 2: p k = p C k 3: else 4: if p U k k then 5: p k = pk C 6: else g k 7: pk B = B 1 k 8: if pk B k then 9: p k = pk B 10: else 11: Solve quadratic eqn p(τ) 2 2 = 0 12: end if 13: end if 14: end if Trust region methods: global/local convergence, approximate January methods 24, 2014 12 / 15

Generalization of dogleg: 2D minimization Suppose: B 0. Search for p(η 1, η 2 ) = η 1 p U + η 2 p B, p. 2D subspace includes p C and Newton step p B = Provably globally convergent Superlinearly locally convergent m(η 1, η 2 ) = m(p(η 1, η 2 )) Trust region methods: global/local convergence, approximate January methods 24, 2014 13 / 15

2D minimization: solving the subproblem Suppose: B 0. Min attained in the interior of the trust-region η m(η 1, η 2 ) = 0 If not = min attained on the boundary. 1D-minimization over parametrized ellipse η 2 1 p U 2 + η 2 2 p B 2 + 2η 1 η 2 (p U ) T p B = 2. Trust region methods: global/local convergence, approximate January methods 24, 2014 14 / 15

Generalization of dogleg: 2D minimization Suppose: B 0 [Byrd, Schnabel, Schultz]. Estimate the smallest eig λ 1 (B) and put α λ 1, α > λ 1. p B = (B + αi ) 1 g If p B > solve 2D minimization problem over span( g, p B ) precisely as in the case B 0. If p B compute eigenvector v 1. Put p = p B + ξv 1, where ξ: p = and ξv T 1 p B 0. (Read hard case, pp. 87 88). Trust region methods: global/local convergence, approximate January methods 24, 2014 15 / 15