Sei sulla pagina 1di 43

department of mathematics and computer science

1
6BV04
DOE: Optimization
Response Surface Methods
department of mathematics and computer science

2
Contents

Optimisation steps

Bo method

Steepest ascent method

!ractica" eamp"e

Response surface desi#ns

Mu"tip"e responses

EVO!

Soft$are

%iterature
department of mathematics and computer science

3
Optimisation steps
Optimisation is achie&ed '( #oin# throu#h the
fo""o$in# phases:

screening )determine $hich factors rea""( in*uence


the outcome+ too": screenin# desi#ns "i,e fractiona"
factoria"-

improvement )approach optimum '( repeated


chan#e of factor settin#s+ too"s: Bo.simp"e or
steepest ascent approach-

determination of optimum )/nd optima" settin#s of


factor settin#s+ too": response surface desi#ns "i,e
CCD or Bo0Behn,en 1 ana"(sis of response surface
usin# ei#en&a"ues-
department of mathematics and computer science

4
current settings
improvement
optimum
department of mathematics and computer science

5
Re#ression mode"s used in optimisation
Statistica" techni2ues for optimisation assume the fo""o$in# )often
reasona'"( satis/ed in practice-:
34ar a$a(5 from the optimum a /rst order mode" often su6ces7
for eamp"e:
Y 8 9
0
1 9
:
x
:
1 9
;
x
;
1
3<ear5 the optimum often a 2uadratic )second order- mode"
su6ces7 4or eamp"e:
Y 8 9
0
1 9
:
x
:
1 9
;
x
;
1 9
:;
x
:
x
;
1

9
::
x
:
;
1 9
;;
x
;
;
1
%ac,0of0/t techni2ues must 'e app"ied in order to chec, $hether
these mode"s are appropriate= since $e cannot direct"( see
$hether $e are near the optimum )cf7 net s"ides-7
department of mathematics and computer science

6
Mode"s
Far away from
optimum:
first order model
department of mathematics and computer science

7
Mode"s
Near optimum:
fitting a first
order model
shows lac!of!
fit
"curvature#
department of mathematics and computer science

$
Mode"s
Near optimum:
second order
model
department of mathematics and computer science

%
>mpro&ement
>n order to e6cient"( mo&e from current factor settin#s to
factor settin# that (ie"d near0optima" &a"ues= ; methods are
a&ai"a'"e:
Bo.Simp"e method
? idea: form ne$ fu"" factoria"s in direction of "ar#est increase
in current fu"" factoria"
? simp"e+ no statistics needed for imp"ementation
? not e6cient
Steepest ascent.descent method
? idea: use :
st
order re#ression mode" from fractiona" factoria"
to o'tain direction of "ar#est increase )3steepest ascent5-
? perform sin#"e runs in direction of "ar#est increase unti"
increase stops
? ad&anced
? recommended since it is more e6cient
department of mathematics and computer science

1&
Bo method
4&'6 41'%
41'$
41'2
41'3
3%'3 4&'%
41'5 4&'&
direction of
largest increase
direction of
largest increase
stop if one has
to return to
previous
settings
department of mathematics and computer science

11
Steepest ascent method
direction of
steepest ascent
contour lines of
first!order model
perpendicular
to contour line
region where
1
e
order!model
has (een determined
department of mathematics and computer science

12
start
screenin#
RSM desi#n
)CCD= 777-
sin#"e o'ser&ation
in direction
steepest ascent
fu"" factoria"
1 centre points
:
st
order
mode" O@A
'etter
o'ser&ationA
stationar(
point
optimumA
stationar(
point
near'(A
#o to
stationar( point
(es
no
(es
no
(es
no
no
accept
stationar( point
end
/t ;
nd
order mode"
(es
Optimization scheme
department of mathematics and computer science

13
!ractica" eamp"e
goal: maimise (ie"d of chemica" reactor
signifcant factors obtained after screening
experiment:

reaction time

reaction temperature
current factor setting: time 8 BC min7 temp 8 :CC DC
current yield: 40 E
department of mathematics and computer science

14
Steepest ascent
;
;
0desi#n $ith C centre points:
time: B0 0 40 min+ temp: :C0 0 :60 DC
resu"ts: mont#omer(:40:7sf

there is no si#ni/cant interaction

there is no si#ni/cant "ac,0of0/t

the re#ression mode" is si#ni/cant


Fence= $e are not near the optimum'
department of mathematics and computer science

15
Steepest ascent path
outcome ana"(sis of measurement:
(ie"d 8 ;47G4 1 07:CCHtime 1 0706CHtemp
$ith coding:
x
:
8 )time0BC-.C x
;
8 )temp0:CC-.C
(ie"d 8 40744 1 07IICHx
:
1 07B;CHx
;
&'775
&'325



direction path: norma" &ector
step size: C min reaction time )choice of chemica" en#ineerJ-
5 &'325
&'42
5 &'775
= coded step size temp )8 ;7:)*#
department of mathematics and computer science

16
Steepest ascent path eperiments
Further e+periments with factor settings of e+periment nr' 1&'
department of mathematics and computer science

17
<ear the optimum
Settin#s eperiment :0:

time 8 KC min

temperature 8 :IC DC
L ;
;
desi#n $ith C centre points is eecuted7
resu"ts: mont#omer(:4047sf

%ac,0of0/t indicates cur&ature7 Fence= $e no$ are
pro'a'"( near the optimum7
department of mathematics and computer science

1$
Muadratic mode"s
>n order to /t a 2uadratic mode" )suita'"e $hen $e
are near the optimum-= $e must &ar( the factors at
3 "e&e"s7
L ;
p
0desi#n $ith centre points does not su6ce=
'ecause then all quadratic factors are confounded7
L B
p
0desi#n is possi'"e= 'ut not to 'e recommended:

num'er of runs #ro$s fast

uses more runs than necessar( to /t 2uadratic


mode"7
department of mathematics and computer science

1%
Response surface desi#ns
Nhe fo""o$in# desi#ns are $ide"( used for /ttin# a
2uadratic mode":

Centra" Composite Desi#n )uniform precision of eOect


estimates-

Bo0Behn,en Desi#n )a"most uniform precision of eOect


estimates= 'ut usua""( fe$er runs re2uired than for CCD-
Nhe choice 'et$een these mode"s is usua""(
decided '( the a&ai"a'i"it( of these desi#ns for a
#i&en num'er of runs and num'er of factors7
<ote that there are other suita'"e desi#ns )usua""(
a&ai"a'"e in statistica" soft$are that supports DOE-7
department of mathematics and computer science

2&
Centra" Composite Desi#n
L CCD consists of B parts:
,
factoria" points

aia" points
,
centre points
L CCD is often eecuted '( addin#
points to an a"read( performed
;
p
0desi#n )hi#h"( e6cient= 'ut 'e$are
of '"oc,in#J-7
department of mathematics and computer science

21
Rotata'i"it(
>n a CCD there are ; possi'"e choices:
number of centre points
location axial points
B( choosin# the aia" points at the "ocations )=0=P=0- etc7
$ith 8 )Q factoria" points-
R
= the desi#n 'ecomes
rotatable= i7e7 the precision )&ariance- of the mode"
depends on the distance to the ori#in on"(7 >n other $ords=
one has the same precision for a"" factor estimates7
department of mathematics and computer science

22
Bo0Behn,en desi#ns
Nhese are desi#ns that
consists of
com'inations from ;
p
0desi#ns7
!roperties:

efcient )fe$ runs-

(almost) rotatable

no corner points of
hypercube
)these are etreme
conditions
$hich are often hard to set-
department of mathematics and computer science

23
Stationar( point
<ear the optimum usua""( a 2uadratic mode"
su6ces:
+ + + + =

<
= =
j i
j i
ij i
k
i
ii i
k
i
i
x x x x Y
2
1 1
&
Fo$ do /nd the optimum after $e correct"(
estimated the parameters usin# a response
surface desi#n )CCD or Bo0Behn,en-A
Nhe net s"ides sho$ the too"s to deri&e optima"
settin#s and the pitfa""s that ha&e to 'e a&oided7
department of mathematics and computer science

24
Recap: optimisation in dimension:
necessar( condition for etremum: :
st
deri&ati&e 8 0
not su6cient:
3point of in*ection5
etra su6cient condition: ;
nd
deri&ati&e 0
department of mathematics and computer science

25
Sero /rst deri&ati&es: sadd"epoint &s7
maimum
saddle point "unfavoura(le# ma+imum "favoura(le#
( )
2 2
x y +
2 2
x y
department of mathematics and computer science

26
Determination of t(pe of optimum
Traphica""(: ma,e contourp"ot )if ;
factors-
Lna"(tica""(:
j i
j i
ij i
k
i
ii i
k
i
i
x x x x Y

<

=

+ + + =
2
1 1
&
matri notation: x B x b x Y
T T
+ + =

&

<ote: B must 'e chosen


as s(mmetric matri=
see eamp"e:
11 12 13
12 22 23
13 23 33
- 2 - 2
- 2 - 2
- 2 - 2














department of mathematics and computer science

27
Stationarit( and matri ana"(sis
stationar( point )zero /rst0order deri&ati&es-:
b B x Bx b
x
Y
1
&
2
1
& 2

= = + =

characterisation throu#h ei#en&a"ues of matri B:


x Bx =

a"" ei#en&a"ues positi&e: min

a"" ei#en&a"ues ne#ati&e: ma

ei#en&a"ues diOerent si#ns: sadd"e point


)the Us are sometimes ca""ed 3parameters of canonica" form5-
department of mathematics and computer science

2$
Stationarit( and matri ana"(sis
>n StatTraphics:

au#ment desi#n

add star points

!"ease note that additiona" centre points are added


and a '"oc, &aria'"e7

Ve can remo&e the centre points from the data set


and i#nore the '"oc, &aria'"e in the ana"(sis7
StatTraphics resu"ts: mont#omer(:4067sf
department of mathematics and computer science

2%
Stationarit( and matri ana"(sis
Wse Mat"a' to a&oid manua" computations:

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> B = [-2.75247 0.5/2 ; 0.5/2 -2.00253] / 2
Analysis Summary
----------------
File name: D:!yD"#s2DS0$#"lle%es&ee'sm"n'%"mery$4-(.s)*
+"mmen': !"n'%"mery 'a,le $4-4
-s'ima'e. e))e#'s )"r "/,ren%s'
----------------------------------------------------------------------
a0era%e = 71.14 2/- 0.$$31(
A:'i4. = $.13114 2/- 0.$33012
B:'em/era'uur = $.03033 2/- 0.$33013
AA = -2.75247 2/- 0.20$705
AB = 0.5 2/- 0.2((003
BB = -2.00253 2/- 0.20$7$
----------------------------------------------------------------------
S'an.ar. err"rs are ,ase. "n '"'al err"r 5i'& 7 ..).
department of mathematics and computer science

3&
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> B = [-2.75247 0.5/2 ; 0.5/2 -2.00253] / 2
B =
-$.37(2 0.$250
0.$250 -$.00$3
department of mathematics and computer science

31
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> ei%6B7
department of mathematics and computer science

32
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> ei%6B7
ans =
-$.4$4$
-0.1(34
'oth ne#ati&e X maimum
department of mathematics and computer science

33
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> , = [$.13114 ; $.03033] /2
Analysis Summary
----------------
File name: D:!yD"#s2DS0$#"lle%es&ee'sm"n'%"mery$4-(.s)*
+"mmen': !"n'%"mery 'a,le $4-4
-s'ima'e. e))e#'s )"r "/,ren%s'
----------------------------------------------------------------------
a0era%e = 71.14 2/- 0.$$31(
A:'i4. = $.13114 2/- 0.$33012
B:'em/era'uur = $.03033 2/- 0.$33013
AA = -2.75247 2/- 0.20$705
AB = 0.5 2/- 0.2((003
BB = -2.00253 2/- 0.20$7$
----------------------------------------------------------------------
S'an.ar. err"rs are ,ase. "n '"'al err"r 5i'& 7 ..).
department of mathematics and computer science

34
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> , = [$.13114 ; $.03033] /2
, =
0.1150
0.5$52
department of mathematics and computer science

35
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> s/#".e. = -0.5 8 in06B7 8 ,
department of mathematics and computer science

36
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> s/#".e. = -0.5 8 in06B7 8 ,
s/#".e. =
0.3313
0.3051
Y :74:4 )distance star point-
X inside eperimenta" re#ion
department of mathematics and computer science

37
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> s/"ri%inal = s/#".e. .8 [5 ; 5] 2 [35 ; $75]
1 2
1 2
$5 175
.
5 5
5 $5. 5 175
time temp
x x
time x temp x

= =
= + = +
Stationarit( and matri ana"(sis
department of mathematics and computer science

3$
Stationarit( and matri ana"(sis
>n Mat"a':

create matri B and &ector '

compute ei#en&a"ues and "ocation of stationar( point


>> s/"ri%inal = s/#".e. .8 [5 ; 5] 2 [35 ; $75]
s/"ri%inal =
3(.14(3
$7(.5213
department of mathematics and computer science

3%
start
screenin#
RSM desi#n
)CCD= 777-
sin#"e o'ser&ation
in direction
steepest ascent
fu"" factoria"
1 centre points
:
st
order
mode" O@A
'etter
o'ser&ationA
stationar(
point
optimumA
stationar(
point
near'(A
#o to
stationar( point
(es
no
(es
no
(es
no
no
accept
stationar( point
end
/t ;
nd
order mode"
(es
Optimization scheme
department of mathematics and computer science

4&
Mu"tip"e responses
>f more than : response &aria'"e needs to 'e
optimised= then a #raphica" $a( of optimisin# ma(
'e achie&ed '( o&er"a(in# contour p"ots in case
there are on"( ; independent &aria'"es7
Overlay Plot
molecular_weight
opbrengst
viscositeit
tijd
t
e
m
p
e
r
a
t
u
u
r
80 82 84 86 88 90
!0
!2
!4
!6
!8
80
department of mathematics and computer science

41
E&o"utionar( Operation )EVO!-
Optimisation of a runnin# production process is not
a"$a(s possi'"e or ma( not 'e a""o$ed 'ecause of costs:

in&o"&es interruption

ma( )temporari"(- (ie"d "o$ 2ua"it( products


Ln a"ternati&e is E&o"utionar( Operation:

eperimentation $ithin runnin# operation

fre2uent eecution of ;
k
0desi#ns= startin# at current
settin#s

hi#h and "o$ settin# of factors are c"ose to each other=


thus no ris, of "o$ 2ua"it( products
department of mathematics and computer science

42
Soft$are

Stat%a' optimisation:
http:..$$$7$in7tue7n".stat"a'
>nteracti&e soft$are for teachin# DOE throu#h cases

Bo: http:..$$$7$in7tue7n".Zmar,o.'o.'o7htm"
Tame0"i,e demonstration of Bo method

Mat"a' &irtua" reactor: Statistics too"'o 0[ Demos 0[


Empirica" Mode"in# 0[ RSM demo

Stat#raphics:
menu choice Specia" 0[ Eperimenta" Desi#n
?
desi#n eperiment $ith pre0de/ned cata"o#ue
?
ana"(sis of eperiments $ith L<OVL
department of mathematics and computer science

43
%iterature

\7 Nr(## and S7 Vo"d7 >ntroduction to Eperimenta"


Desi#n ? Vhat is itA Vh( and Vhere is it Wsefu"A=
Fomepa#e of Chemometrics= editoria" Lu#ust ;00;:
http:..$$$7acc7umu7se.Ztn,]t#.Chemometrics.editori
a".au#;00;7htm"

DOE 'oo,"et from Wmetrics: http:..$$$7umetrics7com


.pdfs.'oo,s.DOE'oo,7pdf

>ntroduction to DOE from moresteam7com http:..


$$$7moresteam7com.too"'o.t40K7cfm

StatSoft E"ectronic Statistics Net'oo,= chapter on


eperimenta" desi#n

<>SN En#ineerin# Statistics Fand'oo,:


http:..$$$7it"7nist7#o&.di&KGK.hand'oo,.

Potrebbero piacerti anche