trust region method

18
Trust Region Methods Part I Andrew R. Conn [email protected] Mathematical Sciences IBM T.J. Watson Research Center February 2007, Montreal 1

Upload: trungnguyenbk

Post on 10-Apr-2018

221 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 1/18

Trust Region MethodsPart I

Andrew R. [email protected]

Mathematical SciencesIBM T.J. Watson Research Center

February 2007, Montreal

1

Page 2: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 2/18

Overview

1 Trust-Region/Modelling Methods

Page 3: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 3/18

Trust-Region/Modelling Methods

What is a trust-region method?

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

Problem: Minimize

−10x 21 + 10x 22 + 4 sinx 1x 2

− 2x 1 + x 

41

3

Page 4: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 4/18

Trust-Region/Modelling Methods

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 0

T R i /M d lli M h d

Page 5: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 5/18

Trust-Region/Modelling Methods

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 1

T t R i /M d lli M th d

Page 6: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 6/18

Trust-Region/Modelling Methods

What is a trust-region method?

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 2

6

Trust Region/Modelling Methods

Page 7: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 7/18

Trust-Region/Modelling Methods

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 3 = x 2

Trust Region/Modelling Methods

Page 8: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 8/18

Trust-Region/Modelling Methods

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 4

8

Trust-Region/Modelling Methods

Page 9: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 9/18

Trust-Region/Modelling Methods

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 5

9

Trust-Region/Modelling Methods

Page 10: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 10/18

Trust Region/Modelling Methods

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

The model and trust region around x 6 (x 7 = x ∗)

10

Trust-Region/Modelling Methods

Page 11: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 11/18

st g o / o g t o s

What is a trust-region method? (continued)

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

-3 -2 -1 0 1 2 3

-3

-2

-1

0

1

2

3

Six iterations —different initial point

11

Trust-Region/Modelling Methods

Page 12: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 12/18

g / g

Basic Algorithm

Initialize: x 0, ∆

Compute Model: mk ( )

Compute Step: Compute s k  from

mins ≤∆

mk (x k  + s )

Trust-region Update: ρ = f  (x k )−f  (x k +s k )mk (x k )−mk (x k +s k )

If  ρ > 0.75 ∆ ← 2.0∆

If 0.25 < ρ < 0.75 ∆ ← ∆

If  ρ < 0.25 ∆ ← 0.5∆

Accept x k  + s k 

Accept x k  + s k 

Reject x k  + s k 

12

Trust-Region/Modelling Methods

Page 13: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 13/18

/

Assumptions to prove convergence

On the problem

Smooth f   ∈ C 

2

Bounded Below f   bounded below

Bounded Hessian xx f   bounded above

13

Trust-Region/Modelling Methods

Page 14: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 14/18

Assumptions to prove convergence

On the model (∀k )

Smooth mk  ∈ C 2

Interpolatesm

k (x k ) =

f  (x k )

Interpolates Gradient x mk (x k ) = x f  (x k )

Bounded Hessianmaxx ∈Bk 

xx mk (x ) bounded above where,Bk  = {x  ∈ n | x  − x k k  ≤ ∆k }

13

Page 15: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 15/18

Page 16: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 16/18

Trust-Region/Modelling Methods

Page 17: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 17/18

Standard Model in the differentiable case

Typical trust region or line search method builds linear or

quadratic model of the objective function f  .The model has to satisfy Taylor-like error bounds.Second Order

|f  (x ) − m(x )| ≤ O(∆3)

|f  

(x 

) − m

(x 

)| ≤ O(∆2

)|2

f  (x ) − 2m(x )| ≤ O(∆)

In fact it typically is a first (or second) order Taylor series

approximation.In derivative based methods constants in O depend only on f  

(and its derivatives).

By reducing the trust region or step size one guarantees better

accuracy.14

Trust-Region/Modelling Methods

Page 18: Trust Region Method

8/8/2019 Trust Region Method

http://slidepdf.com/reader/full/trust-region-method 18/18

Standard Model in the differentiable case

Typical trust region or line search method builds linear or

quadratic model of the objective function f  .The model has to satisfy Taylor-like error bounds.Second Order

|f  (x ) − m(x )| ≤ O(∆3)

|f  

(x 

) − m

(x 

)| ≤ O(∆2

)|2

f  (x ) − 2m(x )| ≤ O(∆)

In fact it typically is a first (or second) order Taylor series

approximation.In derivative based methods constants in O depend only on f  

(and its derivatives).

By reducing the trust region or step size one guarantees better

accuracy.14