34
Engineering Optimization  Concepts and Applications Engineering Optimization Concepts and Applications Fred van Keulen Matthijs Langelaar CLA H21.1  A.vanKeulen@tudelft.nl

ethz_lecture6

Embed Size (px)

Citation preview

Page 1: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 1/34

Engineering Optimization – Concepts and Applications

Engineering Optimization

Concepts and Applications

Fred van KeulenMatthijs Langelaar 

CLA H21.1

 [email protected]

Page 2: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 2/34

Engineering Optimization – Concepts and Applications

Recap / overview

Optimization problem

Definition

Checking

Negative null form Model

Special topics

Sensitivity analysis

Topology optimization

Linear / convex problems

Solution methods

Optimality criteria

Optimization algorithms

Unconstrained problems

Optimality criteria

Optimization algorithms

Constrained problems

Page 3: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 3/34

Engineering Optimization – Concepts and Applications

Summary optimality conditions

● Conditions for local minimum of unconstrained problem:

 – First Order Necessity Condition:

 – Second Order Sufficiency Condition: H positive definite

0 f  

● For convex  f  in convex feasible domain:

condition for global minimum:

 – Sufficiency Condition: 0 f  

Page 4: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 4/34Engineering Optimization – Concepts and Applications

Stationary point nature summary

Definiteness H Nature x*

Positive d. Minimum

Positive semi-d. Valley

Indefinite Saddlepoint

Negative semi-d. Ridge

Negative d. Maximum

HyyT 

0

0

0

0

0

Page 5: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 5/34Engineering Optimization – Concepts and Applications

Complex eigenvalues?

● Question: what is the nature of a stationary point when

H has complex eigenvalues?

●  Answer: this situation never occurs, because H is

symmetric by definition. Symmetric matrices have real

eigenvalues (spectral theory).

Page 6: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 6/34Engineering Optimization – Concepts and Applications

Nature of stationary points

● Nature of initial position depends on load (buckling):

 F  

k 1 

k 2 

l  

5.9,10

2,6

21

k k 

l  F 

21 coscos   l l dz 

 Fdz k k  2

222

1121

21   

0

1222

2111

cossin

cossin

   

   

 Fl k 

 Fl k 

21221

21211

coscossinsin

sinsincoscos

    

    

 Fl k  Fl 

 Fl  Fl k 

0

 Fl k 

 Fl k 

2

1

0

0

 

  

 

k  F crit 

21 ,min

75.4 crit  F 

Page 7: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 7/34Engineering Optimization – Concepts and Applications

Nature of stationary points (2)

75.3 F  75.4 F  75.5 F  75.6 F  75.7 F 

Page 8: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 8/34Engineering Optimization – Concepts and Applications

Unconstrained optimization

algorithms● Single-variable methods

 – 0th order (involving only f ) 

 – 1st order (involving f and f ’ )

 – 2nd order (involving f , f ’ and f ” )

● Multiple variable methods

Page 9: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 9/34

Engineering Optimization – Concepts and Applications

Why optimization algorithms?

● Optimality conditions often cannot be used:

 – Function not explicitly known (e.g. simulation)

 – Conditions cannot be solved analytically

● Example: 122

21

2

1̀21

 x xe xe x x x x f  

Stationary points:

021

021

12

12

221

2

2

211

1

 x x

 x x

e xe x x x

 f  

e x xe x x f  

 f   0

Page 10: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 10/34

Engineering Optimization – Concepts and Applications

0th order methods: pro/con

● Strengths:

 – No derivatives needed

 – Work also for 

discontinuous / non-

differentiable functions

 – Easy to program

 – Robust

● Weaknesses:

 – (Usually) less efficient

than higher order 

methods (many function

evaluations)

Page 11: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 11/34

Engineering Optimization – Concepts and Applications

Minimization with one variable

● Why?

 – Simplest case: good starting point

 – Used in multi-variable methods during line search 

 x x xt  s

 x f   x

..

)(min● Setting:

 f  

 x 

Model

Optimizer 

Iterative process:

Page 12: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 12/34

Engineering Optimization – Concepts and Applications

Termination criteria

● Stop optimization iterations when:

 – Solution is sufficiently accurate (check optimality criteria)

 – Progress becomes too slow:

 – Maximum resources have been spent

 – The solution diverges

 – Cycling occurs

 f  k k  xk k  x f   x f   x x    )()(, 11

 xa  xb 

Page 13: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 13/34

Engineering Optimization – Concepts and Applications

Brute-force approach

● Simple approach: exhaustive search

● Disadvantage: rather inefficient

 f  

 x 

 L0

0

2

1n L L

n

n points:

Final interval size = Ln

Page 14: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 14/34

Engineering Optimization – Concepts and Applications

Basic strategy of 0th order 

methods for single-variable case1. Find interval [a0, b0] that contains the minimum

(bracketing)

2. Iteratively reduce the size of the interval [ak , b

k ] 

(sectioning)

3.  Approximate the minimum by the minimum of a simple

interpolation function over the interval [a N , b N ] 

● Sectioning methods:

 – Dichotomous search

 – Fibonacci method

 – Golden section method

Page 15: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 15/34

Engineering Optimization – Concepts and Applications

Bracketing the minimum

 f  

 x 

 x1   x2 = x1+  x3 = x2+g  x4 = x3+g2

[a0, b0]

Starting point x1, stepsize , expansion parameter g: user-defined

Page 16: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 16/34

Engineering Optimization – Concepts and Applications

Unimodality

● Bracketing and sectioning methods work best for 

unimodal functions:

“An unimodal function consists of exactly one

monotonically increasing and decreasing part”  

Page 17: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 17/34

Engineering Optimization – Concepts and Applications

 L0 

a0  b0 

Dichotomous searchMain Entry: di·chot·o·mous Pronunciation: dI-'kät-&-m&s also d&-

Function: adjective 

: dividing into two parts

● Conceptually simple

idea:

 – Try to split interval in half in each step

d << L0  L0 /2

: f   f  

Page 18: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 18/34

Engineering Optimization – Concepts and Applications

Dichotomous search (2)

● Interval size after 1 step (2 evaluations):

d  012

1 L L

● Interval size after m steps (2m evaluations):

 

  

 

mmm

 L L

2

11

2

0 d 

● Proper choice for d :

m

mm

mm

 L L L L L

21010102

0

ideal

0ideal

 L0 

Page 19: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 19/34

Engineering Optimization – Concepts and Applications

Dichotomous search (3)

● Example: m = 10

10242

0

10

0ideal

10

 L L L

1024010

0

ideal

10 L L d 

 

  

 

0

10 log L

 Lm

Idealinterval

reduction

10240

0 Ld 

Page 20: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 20/34

Engineering Optimization – Concepts and Applications

Sectioning - Fibonacci

● Situation:

minimum

bracketed

between x1 

and x3 : x1   x3 

Fibonacci,

1180?-1250?

 x2 

● Test new points and reduce interval

 x4  x4 

● Optimal point placement?

Page 21: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 21/34

Engineering Optimization – Concepts and Applications

Optimal sectioning

● Fibonacci method: optimal sectioning method

● Given:

 – Initial interval [a0, b0]

 – Predefined total number of evaluations N, or: 

 – Desired final interval size  

Page 22: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 22/34

Engineering Optimization – Concepts and Applications

Fibonacci sectioning - basic idea

● Start at final interval and use symmetry and maximum

interval reduction:

 I  N-1 = 2 I  N 

 I  N-2 = 3 I  N 

 I  N-3 = 5 I  N 

 I  N-4 = 8 I  N 

 I  N-5 = 13 I  N 

d << I  N  

 I  N  

 N  j j N  I  F  I  1 Fibonacci number  j F 

12 k k k  I  I  I  j N  j N  j N  I  I  I  12

Page 23: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 23/34

Engineering Optimization – Concepts and Applications

Sectioning – Golden Section

● For large N , Fibonacci fraction b converges to golden

section ratio f  (0.618034…): 

 N 

 N 

 F 

 F 1

● Golden section method

uses this constant interval

reduction ratio f 

f  

f  1 

Page 24: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 24/34

Engineering Optimization – Concepts and Applications

Sectioning - Golden Section

● Origin of golden section: I 1 

 I 3 = f  I 2

 I 2 = f  I 1

 I 2 = f  I 1321 I  I  I 

1

2

11 I  I  I  f f 

2

5101 2,1

2 f f f 

618034.02

15

● Final interval: 1 I  I  N 

 N  f 

Page 25: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 25/34

Engineering Optimization – Concepts and Applications

 

  

 

0

10 log L

 Lm

Ideal dichotomous

interval reduction

Fibonacci

Golden

section

Evaluations

Comparison sectioning methods

 N  

Dichotomous 12

Golden section 9

Fibonacci 8

(Exhaustive 99)

Example:

reduction to 2% of 

original interval:

● Conclusion: Golden section simple and near-optimal

Page 26: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 26/34

Engineering Optimization – Concepts and Applications

Quadratic interpolation

● Three points of the bracket define interpolating quadratic

function:

cbxax x f  

2

)(

~

ai  bi ● New point evaluated at

minimum of parabola:a

b xbax f   new

202'

~

 xnew 

● For minimum: a > 0!

● Shift xnew when very close to existing point

ai+1  bi+1 

Page 27: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 27/34

Engineering Optimization – Concepts and Applications

Unconstrained optimization

algorithms● Single-variable methods

 – 0th order (involving only f )

 – 1st order (involving f and f ’ ) 

 – 2nd order (involving f , f ’ and f ” )

● Multiple variable methods

Page 28: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 28/34

Engineering Optimization – Concepts and Applications

Cubic interpolation

● Similar to quadratic interpolation, but with 2 points and

derivative information:

d cxbxax x f   23

)(

~

ai  bi 

Page 29: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 29/34

Engineering Optimization – Concepts and Applications

Bisection method

● Optimality conditions: minimum at stationary point

Root finding of  f ’ 

● Similar to sectioning methods, but uses derivative:

 f    f ’  

Page 30: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 30/34

Engineering Optimization – Concepts and Applications

Secant method

●  Also based on root finding of  f ’ 

● Uses linear interpolation

 f ’ 

Page 31: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 31/34

Engineering Optimization – Concepts and Applications

Unconstrained optimization

algorithms● Single-variable methods

 – 0th order (involving only f )

 – 1st order (involving f and f ’ )

 – 2nd order (involving f , f ’ and f ” ) 

● Multiple variable methods

Page 32: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 32/34

Engineering Optimization – Concepts and Applications

Newton’s method 

●  Again, root finding of  f ’ 

● Basis: Taylor approximation of  f ’:2

'( ) '( ) ''( ) ( ) f x h f x f x h o h Linear approximation

)("

)('

 x f  

 x f  h

● New guess:)(")('

1

k k k k k 

 x f   x f   xh x x

Page 33: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 33/34

Engineering Optimization – Concepts and Applications

Newton’s method 

● Best convergence of all methods:

● Unless it diverges

 xk  

 f’  

 xk+1 

 xk+2 

 f’  

 xk  

 xk+1   xk+2 

Page 34: ethz_lecture6

7/30/2019 ethz_lecture6

http://slidepdf.com/reader/full/ethzlecture6 34/34

Engineering Optimization – Concepts and Applications

Summary single variable methods

● Bracketing +

Dichotomous sectioning

Fibonacci sectioning

Golden ratio sectioning

Quadratic interpolation

Cubic interpolation

Bisection method

Secant method

Newton method

●  And many, many more!

0th order 

1st order 

2nd order 

In practice:additional “tricks”needed to dealwith:

Multimodality

Strongfluctuations

Round-off errors

Divergence