Professional Documents
Culture Documents
May 4, 2006
Nikolaus Hansen
Meta-Parameters
how many parameters of the algorithm need to be adjusted
to the object function?
how many different setups were tested?
how many different setups were finally used?
Nikolaus Hansen
Methods
task: black-box optimization of 25 benchmark functions
25 runs on each benchmark function for each dimension
n = 10, 30
a run is successful if the global optimum is reached with
the given precision, before the
maximumnumber of function evaluations
105 for n = 10
FEmax =
is reached
3 105 for n = 30
Remark
the setting of FEmax has a remarkable influence on the results, if the target function value can be
reached only for a (slightly) larger number of function evaluations with a high probability. Where
FEs FEmax the result must be taken with great care.
Reference
Suganthan, Hansen, Liang, Deb, Chen, Auger, and Tiwari (2005). Problem Definitions and Evaluation
Criteria for the CEC 2005 Special Session on Real-Parameter Optimization, Technical report, Nanyang
Technological University, Singapore, May 2005, http://www.ntu.edu.sg/home/EPNSugan
Nikolaus Hansen
References to Algorithms
BLX-GL50
BLX-MA
CoEVO
DE
DMS-L-PSO
EDA
G-CMA-ES
K-PCX
L-CMA-ES
L-SaDE
SPC-PNX
Ronkk
onen
et al. (Real-Parameter Optimization. . . )
Liang and Suganthan (Dynamic Multi-Swarm. . . )
Yuan and Gallagher (Experimental Results. . . )
Auger and Hansen (A Restart CMA. . . )
Sinha et al. (A Population-Based,. . . )
Auger and Hansen (Performance Evaluation. . . )
Qin and Suganthan (Self-Adaptive Differential. . . )
Ballester et al. (Real-Parameter Optimization. . . )
In: CEC 2005 IEEE Congress on Evolutionary Computation, Proceedings
Nikolaus Hansen
Summarized Results
Empirical Distribution of Normalized Success Performance
n = 10
0.9
DE
0.8
LSaDE
DMSLPSO
0.7
EDA
0.6
BLXGL50
0.5
LCMAES
0.4
KPCX
0.3
SPCPNX
0.2
BLXMA
0.1
CoEVO
0 0
10
10
10
FEs / FEsbest
FEs = mean(#fevals)
GCMAES
0.9
KPCX
0.8
LCMAES
DMSLPSO
0.7
LSaDE
0.6
EDA
0.5
SPCPNX
0.4
BLXGL50
0.3
CoEVO
0.2
BLXMA
0.1
0 0
10
n = 30
GCMAES
DE
1
10
10
FEs / FEsbest
Small values for FEs and therefore large (cumulative frequency) values in the graphs
are preferable.
Nikolaus Hansen
Function Sets
Nikolaus Hansen
Unimodal Functions
Normalized Success Performance, Tabulated
n = 10, FEmax = 100000
solved success
rate
functions
G-CMA-ES
EDA
DE
L-CMA-ES
BLX-GL50
DMS-L-PSO
L-SaDE
SPC-PNX
CoEVO
K-PCX
BLX-MA
6
6
6
6
5
5
5
4
4
4
3
100%
97%
96%
88%
83%
80%
77%
67%
67%
62%
49%
1S
re
phe
1000
1.6(25)
10.0(25)
29.0(25)
1.7(25)
19.0(25)
12.0(25)
10.0(25)
6.7(25)
23.0(25)
1.0(25)
12.0(25)
6
e
ds
10
oun
Nois
ion
with
on B
ndit
2
2
6
.
.
.
o
ck
1
1
2
C
efel
efel
efel
oid
nbro
ose
chw
llips
chw
chw
2S
3E
4S
5S
6R
2400
6500
1.0(25) 1.0(25)
4.6(25) 2.5(23)
19.2(25) 18.5(20)
1.1(25) 1.0(25)
17.1(25)
[9]
5.0(25) 1.8(25)
4.2(25) 8.0(16)
12.9(25)
[11]
11.3(25) 6.8(25)
1.0(25)
[8]
15.4(25)
[10]
2900
5900
1.0(25) 1.0(25)
4.1(25) 4.2(25)
17.9(25) 6.9(25)
65.5( 7) 1.0(25)
14.5(25) 4.7(25)
[11] 18.6(20)
15.9(24)
[9]
10.7(25) 6.8(25)
16.2(25)
[10]
19.7(21)
[11]
25.9(24)
[8]
7100
1.5(25)
9.6(22)
6.6(24)
1.3(25)
7.3(25)
7.7(25)
6.9(25)
[10]
[11]
1.0(22)
[9]
of the best
Table entries: Success Performance FEs divided by FEs of the best algorithm (first
row), (number of successful runs in round brackets), [rank of median final function
value, where FEs is not available, in square brackets]
Nikolaus Hansen
Unimodal Functions
Normalized Success Performance, Tabulated
solved success
functions rate
G-CMA-ES
L-CMA-ES
EDA
DMS-L-PSO
BLX-GL50
SPC-PNX
K-PCX
L-SaDE
BLX-MA
DE
CoEVO
6
5
4
4
3
4
3
3
1
1
2
90%
83%
67%
63%
50%
45%
43%
41%
17%
17%
7%
re
phe
2700
1.7(25)
1.8(25)
55.6(25)
1.9(25)
21.5(25)
11.1(25)
1.0(25)
7.4(25)
11.9(25)
51.9(25)
519 ( 3)
6
ds
oise
10
oun
ith N
nB
ition
.6 o
.2 w
ond
ck
2
1
C
l
l
r
e
e
ef
ef
oid
nb o
ose
llips
chw
chw
3E
4S
5S
6R
.2
1
efel
chw
2S
12000
1.1(25)
1.2(25)
13.3(25)
10.8(25)
13.3(25)
26.7(22)
1.0(25)
12.5(24)
[10]
[11]
70.0( 8)
43000
1.0(25)
1.0(25)
5.1(25)
7.9(21)
[7]
[11]
[6]
[8]
[10]
[9]
[5]
59000
1.0(10)
[11]
3.4(25)
[9]
[6]
6.1(19)
[8]
9.2(13)
[7]
[5]
[10]
66000 60000
1.0(25) 1.0(25)
1.1(25) 1.1(25)
[3]
[10]
[9] 5.5(24)
[6] 3.7(25)
[10] 86.7( 1)
[7] 1.1(14)
[4]
[8]
[7]
[8]
[5]
[7]
[11]
[11]
Table entries: Success Performance FEs divided by FEs of the best algorithm
(first row), (number of successful runs in round brackets), [rank of median final
function value, where FEs is not available, in square brackets]
Nikolaus Hansen
Unimodal Functions
Empirical Distribution of Normalized Success Performance
n = 10
0.9
EDA
0.8
DE
LCMAES
0.7
LSaDE
0.6
DMSLPSO
0.5
BLXGL50
0.4
SPCPNX
0.3
KPCX
0.2
CoEVO
0.1
BLXMA
0 0
10
10
FEs / FEsbest
10
GCMAES
0.9
LCMAES
0.8
DMSLPSO
EDA
0.7
SPCPNX
0.6
KPCX
0.5
LSaDE
0.4
BLXGL50
0.3
CoEVO
0.2
BLXMA
0.1
0 0
10
n = 30
GCMAES
DE
1
10
10
FEs / FEsbest
Empirical distribution function of the Success Performance FEs divided by FEs of the
best algorithm (table entries of last slides).
#all runs (25)
FEs = mean(#fevals) #successful
, where #fevals includes only successful runs.
runs
Small values of FEs and therefore large values in the empirical distribution graphs are
preferable.
Nikolaus Hansen
G-CMA-ES
L-SaDE
DMS-L-PSO
K-PCX
DE
L-CMA-ES
BLX-GL50
BLX-MA
EDA
SPC-PNX
CoEVO
5
4
4
3
5
2
3
2
3
2
0
63%
53%
47%
40%
30%
25%
17%
15%
9%
1%
0%
nk
t
ou
wa
r ie
7G
ds
ble
un
Bo
tr
as
9R
4700
17000
1.0(25) 4.5(19)
36.2( 6) 1.0(25)
126 ( 4) 2.1(25)
[10] 2.9(24)
255 ( 2) 10.6(11)
1.2(25)
[11]
12.3( 9) 10.0( 3)
[11] 5.7(18)
404 ( 1)
[9]
383 ( 1)
[8]
[9]
[10]
str
10
Ra
ed
tat
ara
ep
S
igin
igin
Ro
as
str
ier
e
1W
s
12
fel
we
ch
3
2.1
15
Hy
ep
dS
ara
ble
bri
55000 190000
8200 33000
1.2(23) 1.4( 6) 4.0(22)
[6]
[5]
[8] 3.9(25) 1.0(23)
[3]
[7] 6.6(19) 1.7(22)
1.0(22)
[10] 1.0(14)
[11]
[9] 1.0(12) 8.8(19) 75.8( 1)
[10]
[6] 11.6(12)
[6]
[5]
[5] 12.1(13)
[9]
[7]
[9]
[9] 8.5( 5)
[4] 2.9( 3) 4.3(10)
[9]
[8] 5.8( 1)
[10]
[6]
[11]
[11]
[11]
[8]
#fevals
psucc
Table entries: Success Performance FEs normalized (divided) by the first row,
(number of successful runs in round brackets), [rank of median final function value,
where FEs is not available, in square brackets]
Nikolaus Hansen
K-PCX
G-CMA-ES
L-SaDE
DMS-L-PSO
EDA
BLX-GL50
DE
L-CMA-ES
SPC-PNX
CoEVO
BLX-MA
4
5
2
2
1
1
1
1
1
1
1
38%
37%
36%
22%
20%
20%
20%
20%
13%
9%
7%
nk
wa
r ie
ds
un
Bo
9R
le
rab
pa
trig
as
7G
6100
2.5(10)
1.0(25)
21.3(20)
9.8(24)
21.3(25)
10.2(25)
32.8(22)
1.1(25)
60.7(16)
93.4(11)
[11]
t
ou
e
nS
str
10
Ra
tat
igin
Ro
ed
11
ie
We
ras
rst
ble
ara
.13
ep
l2
e
S
f
e
d
bri
hw
Sc
Hy
12
15
Table entries: Success Performance FEs divided by FEs of the best algorithm
(first row), (number of successful runs in round brackets), [rank of median final
function value, where FEs is not available, in square brackets]
Nikolaus Hansen
Multimodal Functions
n = 10
GCMAES
0.9
DE
0.8
LSaDE
DMSLPSO
0.7
KPCX
0.6
BLXGL50
0.5
EDA
0.4
BLXMA
0.3
LCMAES
0.2
SPCPNX
0.1
CoEVO
0 0
10
10
10
FEs / FEsbest
GCMAES
0.9
KPCX
0.8
DMSLPSO
LSaDE
0.7
LCMAES
0.6
BLXMA
0.5
BLXGL50
0.4
EDA
0.3
DE
0.2
SPCPNX
0.1
0 0
10
CoEVO
1
10
FEs / FEsbest
10
Empirical distribution function of the Success Performance FEs divided by FEs of the
best algorithm (table entries of last slides).
#all runs (25)
FEs = mean(#fevals) #successful
, where #fevals includes only successful runs.
runs
Small values of FEs and therefore large values in the empirical distribution graphs are
preferable.
Nikolaus Hansen
G-CMA-ES
BLX-GL50
L-SaDE
DMS-L-PSO
L-CMA-ES
SPC-PNX
EDA
BLX-MA
K-PCX
DE
CoEVO
mean
3.8
4.2
5.3
5.7
6.0
6.2
6.2
6.6
7.0
7.0
8.2
s
6
2
ion inuou
rF
dit
e
t
10
ffe
ds
on
on
ois
un
w
ion 6&7 Scha ated
o
C
C
t
o
N
o
i
h
B
d
n
t
rr
h
1
4
8
tB
ed
ed
on
F2
Hig d No d F2 d ou
Ro d wit d F1 d Na d on
y C pand pand brid
ri
rid
rid
ri
ri
ri
ri
ri
ri
e
b
b
b
b
b
b
b
b
b
l
ck
Hy 7 Hy 8 Hy 9 Hy 0 Hy 1 Hy 2 Hy 3 Hy 4 Hy 5 Hy
Ex 4 Ex
8A
1
16
1
1
2
2
2
2
2
2
13
1
5.5
5.5
5.5
5.5
5.5
11
5.5
5.5
5.5
5.5
5.5
4
6
1
2
3
8
10
7
5
11
9
7
1.5
6
3.5
11
8
5
1.5
3.5
10
9
1
2.5
5.5
4
7.5
7.5
9
5.5
2.5
11
10
5.5
3
3
3
11
5.5
8
7
1
10
9
3
3
8
8
6
3
3
8
10
3
11
2.5
2.5
7.5
7.5
5
2.5
7.5
7.5
11
2.5
10
2.5
2.5
7.5
7.5
5
2.5
7.5
7.5
10
2.5
11
5.5
5.5
5.5
5.5
1
5.5
5.5
10
11
5.5
5.5
1.5
5.5
5.5
5.5
3
9
8
5.5
1.5
10
11
4
4
4
8
4
4
4
10
11
4
9
5
5
5
5
11
5
5
5
10
5
5
2.5
8.5
4.5
8.5
4.5
8.5
2.5
6
8.5
11
1
Table entries: Rank of the median of the final best function values from 25 runs
measured with two digits of precision
Nikolaus Hansen
G-CMA-ES
EDA
BLX-MA
SPC-PNX
BLX-GL50
L-CMA-ES
DE
K-PCX
CoEVO
L-SaDE
DMS-L-PSO
mean
4.1
4.8
4.9
5.0
5.1
5.6
6.1
6.2
8.5
6
2
ion inuo
rF e
dit
e
10
ffe
ds
on Cont
d
ois
ha arabl
un
w
e
ion 6&7
C
c
t
t
o
N
o
i
S
h
B
d
n
ta
rr
p
h
1
8
4
ed
ed
on
Se d Ro d wit d F1 d Na d on
F2 d Hig d No d F2
y C pand pand brid
ri
ri
ri
ri
ri
rid
ri
ri
ri
y
yb
yb Hyb
yb
yb
yb
yb Hyb
yb
yb
x
x
kle
c
E
H
H
H
H
H
H
H
H
E
H
14
15
16
17
18
19
20
21
22
23
24
25
8A
13
2.5
7.5
7.5
7.5
7.5
2.5
7.5
11
7.5
2.5
2.5
5
11
4
8
6.5
2
6.5
10
9
1
3
6.5
6.5
6.5
6.5
2
10
6.5
10
6.5
2
2
1
8
4.5
8
4.5
2
8
11
10
4.5
4.5
1
7
10
4.5
4.5
3
8
2
9
6
7
5
2
3
10
8
1
9
5.5
1
3
5.5
5.5
8.5
5.5
2
10
8.5
5.5
1
3
5.5
5.5
8.5
5.5
2
10
8.5
5.5
1
3
5.5
5.5
8.5
5.5
2
10
8.5
4.5
4.5
4.5
4.5
4.5
4.5
4.5
9
4.5
1
3
7
3
5
3
6
9
8
3.5
3.5
3.5
3.5
7
3.5
3.5
9
8
6
2.5
2.5
2.5
7
9
2.5
5
8
4
4
4
4
4
4
8
4
9
Table entries: Rank of the median of the final best function values from 25 runs
measured with two digits of precision
Nikolaus Hansen