Boosting Genetic Algorithms With Self-Adaptive Selection
Boosting Genetic Algorithms With Self-Adaptive Selection
I. I NTRODUCTION
Parameter control is a long-standing grand challenge in
evolutionary computing. The motivation behind this interest
is mainly twofold: performance increase and the vision of
parameterless evolutionary algorithms (EA). The traditional
mainstream of research concentrated on adaptive or selfadaptive control of the variation operators, mutation and
recombination [15], [9], [11]. However, there is recent evidence, or at least strong indication, that tweaking other
EA components can be more rewarding. In [2], [7] we have
addressed population size, in the present study we consider
selection.
Our approach to selection is based on a radically new
philosophy. Before presenting the details, let us rst make
a few observations about (self-)adaptation and the control of
selection parameters. Globally, two major forms of setting
parameter values in EAs are distinguished: parameter tuning
and parameter control [6], [8]. By parameter tuning we
mean the commonly practiced approach that amounts to
nding good values for the parameters before the run of
the algorithm and then running the algorithm using these
values, which remain xed during the run. Parameter control
forms an alternative, as it amounts to starting a run with
initial parameter values that are changed during the run.
Further distinction is made based on the manner changing
the value of a parameter is realized (i.e., the how-aspect).
Parameter control mechanisms can be classied into one of
the following three categories.
A.E. Eiben, M.C. Schut and A.R. de Wilde are with the Department
of Computer Science, Vrije Universiteit, Amsterdam (email: {AE.Eiben,
MC.Schut}@few.vu.nl).
0-7803-9487-9/06/$20.00/2006 IEEE
1584
While the picture regarding the control of variation operators in EAs is rather diverse, most researchers agree that
increasing the selection pressure as the evolutionary process
goes on offers advantages [1], [18], [5], [16], [10], [14],
[12]. In the present investigation we will introduce as little as
possible bias towards increasing or decreasing the parameter
values and make selection pressure an observable.
The basis for both mechanisms is an extra parameter k
[kmin , kmax ] in each individual and a voting mechanism
that determines the tournament size K used in the GA. It is
1585
1
1 k N (0,1)
k = k 1 +
e
k
(5)
with = 0.22.
This mechanism differs from pure self-adaptation because of the heuristic rule specifying the direction of the
change. However, it could be argued that this mechanism
is not a clean adaptive scheme (because the initial k values
are inherited), nor a clean self-adaptive scheme (because the
nal k values are inuenced by a user dened heuristic), but
some hybrid form. For this reason we perceive and name this
mechanism hybrid self-adaptive (HSA).
III. E XPERIMENTAL SETUP
A. Test suite
The test suite1 for testing GAs is obtained through the
Multimodal Problem Generator of Spears [17]. We generate
landscapes of 1, 2, 5, 10, 25, 50, 100, 250, 500 and 1000
binary peaks P whose heights are linearly distributed and
where the lowest peak is 0.5. The length L of these bit
strings is 100. The tness of an individual is measured by
the Hamming distance between the individual and the nearest
peak, scaled by the height of that peak. The nearest peak is
determined by
P
TABLE I
D ESCRIPTION OF THE SGA.
1 The
1586
steady-state
100-bitstring
uniform crossover
0.5
bit-ip
0.01
k-tournament
delete-worst-two
100
random
f (x) = 1 or 10,000 evaluations
C. Performance measures
During running a GA, each generation is monitored by
measuring the Best Fitness (BF), Mean Fitness (MF), Worst
Fitness (WF) and the diversity of the population. After 100
runs, data of the monitors is collected and the Mean Best
Fitness (MBF) and its standard deviation (SDMBF), the
Average number of Evaluations to a Solution (AES) and its
standard deviation (SDAES) and the Success Rate (SR) will
be calculated.
Peaks
SR
AES
SDAES
MBF
SDMBF
1
2
5
10
25
50
100
250
500
1000
100
100
100
89
63
45
14
12
7
4
989
969
1007
1075
1134
1194
1263
1217
1541
1503
244
206
233
280
303
215
220
166
446
272
1.0
1.0
1.0
0.9939
0.9879
0.9891
0.9847
0.9850
0.9876
0.9862
0.0
0.0
0.0
0.0175
0.0190
0.0127
0.0140
0.0131
0.0119
0.0113
TABLE III
E ND RESULTS OF GAHSAT.
IV. R ESULTS
SGA
In table II is to see that the harder the problem, the lower
the SR and MBF and the higher the AES.
GAHSAT
In table III the results for GAHSAT are given. These
results are promising because the AES values are much
lower than those of the SGA. This indicates that on-the-y
adjustment of K contributes to a faster GA. Comparing the
SRs and MBF results (and the SDs) of SGA and GAHSAT
makes clear that the extra speed of GAHSAT comes at no
extra costs in terms of solution quality or stability.
Peaks
SR
AES
SDAES
MBF
SDMBF
1
2
5
10
25
50
100
250
500
1000
100
100
100
92
62
46
21
16
3
1
1312
1350
1351
1433
1485
1557
1669
1635
1918
1675
218
214
254
248
280
246
347
336
352
0
1.0
1.0
1.0
0.9956
0.9893
0.9897
0.9853
0.9867
0.9834
0.9838
0.0
0.0
0.0
0.0151
0.0164
0.0128
0.0147
0.0130
0.0146
0.0126
TABLE IV
E ND RESULTS OF GASAT.
GASAT
The outcomes for GASAT in Table IV indicate that the
purely self-adaptive selection is not as powerful as the
adaptive or hybrid self-adaptive mechanism (in table III).
But nevertheless GASAT has better performance than SGA,
reconrming that on-the-y adjustment of K is benecial.
Peaks
SR
AES
SDAES
MBF
SDMBF
1
2
5
10
25
50
100
250
500
1000
100
100
100
93
62
37
22
11
6
5
1478
1454
1488
1529
1674
1668
1822
1923
2089
2358
191
143
159
168
238
221
198
206
230
398
1.0
1.0
1.0
0.9961
0.9885
0.9876
0.9853
0.9847
0.9865
0.9891
0.0
0.0
0.0
0.0142
0.0174
0.0140
0.0145
0.0137
0.0122
0.0100
TABLE II
E ND RESULTS OF SGA.
1587
:
:
x
1 = x
2 ,
x
1 = x
2 ,
x
1 > x
2 ,
Peaks
1
2
5
10
25
50
100
250
500
1000
p
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.23
0.61
p
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.88
0.69
<
<
<
<
<
<
<
<
=
=
V. C ONCLUSIONS
In this paper we have presented experimental evidence
that varying the selection pressure on-the-y can signicantly
improve the performance of a GA. This is, perhaps, not such
a novel result per se. However, it is interesting that here we
use a simple mechanism2 and apply no sophisticated twists
to it. Yet we obtain a GA that compares favorably with the
best GA we found for the same test suite in an earlier paper
[7]. The comparison between the former winner, a GA with
adaptive population size (APGA), and GAHSAT is shown
in Table VI. Note that the MBF results are omitted for they
showed no signicant difference. This comparison shows that
the GAHSAT is very competitive, running out the APGA on
the smoother landscapes.
Note also that the mechanism includes an additional parameter (the learning rate). From the perspective of increasing EA performance this is not a problem. However, from
the parameterless evolutionary algorithms angle this may be
considered problematic, because we eliminate one parameter
(K) at the cost of introducing another one (). Nevertheless,
in general, we believe that meta-parameters, i.e., those of a
learning method extending the EA, are less sensitive for an
accurate setting that the technical parameters, i.e., those of
the EA itself. Furthermore, in our particular case we did not
tune but used the recommended value and found that it
worked well.
TABLE V
t- TEST RESULTS FOR THE FOUND RANKING .
1
K
0
100
200
300
400
Generations
500
600
700
800
GAHSAT
APGA
Peaks
SR
AES
SR
AES
1
2
5
10
25
50
100
250
500
1000
100
100
100
89
63
45
14
12
7
4
989
969
1007
1075
1134
1194
1263
1217
1541
1503
100
100
100
95
54
35
22
8
6
1
1100
1129
1119
1104
1122
1153
1216
1040
1161
910
TABLE VI
C OMPARING GAHSAT AND THE WINNING APGA FROM [7].
1
K
0
Fig. 1.
100
200
300
400
Generations
500
600
700
800
1588
2 that
R EFERENCES
[1] E.H.L. Aarts and J. Korst. Simulated Annealing and Boltzmann
Machines. Wiley, Chichester, UK, 1989.
[2] T. Back, A.E. Eiben, and N.A.L. van der Vaart. An empirical
study on GAs without parameters. In M. Schoenauer, K. Deb,
G. Rudolph, X. Yao, E. Lutton, J.J. Merelo, and H.-P. Schwefel,
editors, Proceedings of the 6th Conference on Parallel Problem Solving
from Nature, number 1917 in Lecture Notes in Computer Science,
pages 315324. Springer, Berlin, 2000.
[3] T. Back, D.B. Fogel, and Z. Michalewicz, editors. Evolutionary
Computation 1: Basic Algorithms and Operators. Institute of Physics
Publishing, Bristol, 2000.
[4] Th. Back and M. Schutz. Intelligent mutation rate control in canonical
genetic algorithms. In Zbigniew W. Ras and Maciej Michalewicz, editors, Foundations of Intelligent Systems, 9th International Symposium,
ISMIS 96, Zakopane, Poland, June 9-13, 1996, Proceedings, volume
1079 of Lecture Notes in Computer Science, pages 158167. Springer,
Berlin, Heidelberg, New York, 1996.
[5] A. Dukkipati, N. M. Musti, and S. Bhatnagar. Cauchy annealing
schedule: An annealing schedule for boltzmann selection scheme in
evolutionary algorithms. pages 5562.
[6] A.E. Eiben, R. Hinterding, and Z. Michalewicz. Parameter control in
evolutionary algorithms. IEEE Transactions on Evolutionary Computation, 3(2):124141, 1999.
[7] A.E. Eiben, E. Marchiori, and V.A. Valko. Evolutionary Algorithms
with on-the-y Population Size Adjustment. In X. Yao et al., editor,
Parallel Problem Solving from Nature, PPSN VIII, number 3242 in
Lecture Notes in Computer Science, pages 4150. Springer, Berlin,
Heidelberg, New York, 2004.
[8] A.E. Eiben and J.E. Smith. Introduction to Evolutionary Computing.
Springer, 2003.
[9] D.B. Fogel. Other selection methods. In Back et al. [3], chapter 27,
pages 201204.
[10] L. Gonzalez and J. Cannady. A self-adaptive negative selection
approach for anomaly detection. In 2004 Congress on Evolutionary
Computation (CEC2004), pages 15611568. IEEE Press, Piscataway,
NJ, 2004.
[11] P.J.B. Hancock. A comparison os selection mechanisms. In Back et al.
[3], chapter 29, pages 212227.
[12] N. Hansen, A. Gawelczyk, and A. Ostermeier. Sizing the population
with respect to the local progress in (1, lambda)-evolution strategies.
In Proceedings of the IEEE Conference on Evolutionary Computation,
pages 8085. IEEE Press, 1995. (Authors in alphabetical order).
[13] G. Hardin. The tradegy of the commons. Science, 162:12431248,
December 1968.
[14] M. Herdy. The number of offspring as a strategy parameters in
hierarchically organized evolution strategies. ACM SIGBIO Newsletter,
13(2):29, 1993.
[15] S.W. Mahfoud. Boltzmann selection. In Back et al. [3], chapter 26,
pages 195200.
[16] E. Poupaert and Y. Deville. Acceptance driven local search and
evolutionary algorithms. In L. Spector, E. Goodman, A. Wu, W.B.
Langdon, H.-M. Voigt, M. Gen, S. Sen, M. Dorigo, S. Pezeshk,
M. Garzon, and E. Burke, editors, Proceedings of the Genetic and
Evolutionary Computation Conference (GECCO-2001), pages 1173
1180. Morgan Kaufmann, 2001.
[17] W.M. Spears. Evolutionary Algorithms: the role of mutation and
recombination. Springer, Berlin, Heidelberg, New York, 2000.
[18] E. Zitzler and S. Kunzli. Indicator-based selection in multiobjective
search. In E.K. Burke X. Yao, J.A. Lozano, J. Smith, J.J.M. Guervos,
J.A. Bullinaria, J.E. Rowe, P. Tino, A. Kaban, and H.-P. Schwefel,
editors, PPSN, number 3242 in Lecture Notes in Computer Science,
pages 832842. Springer, Berlin, Heidelberg, New York, 2004.
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Best Fitness
Mean Fitness
Worst Fitness
Diversity
0.2
0.1
100
200
300
400
Generations
500
600
700
800
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
Best Fitness
Mean Fitness
Worst Fitness
Diversity
0.1
100
200
300
400
Generations
500
600
700
800
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
Best Fitness
Mean Fitness
Worst Fitness
Diversity
0.1
100
200
300
400
Generations
500
600
700
800
1589