Comparison of Different Evolutionary Methodologies

Applied to Electronic Filter Design

Ricardo S. Zebulum Marco Aurelio Pacheco Marley Vellasco

1- Centre for Computational Neurosciences and Robotics, University of Sussex

Brighton B N l 9SB, UK, E-mail : ricardoz@cogs .susx .ac .uk
2 - Centro de Inteligencia Computacionul Aplicada
Pontijicia Universidade Catolica, Rio de Janeiro - Brasil

Abstract 2. Evolutionary Electronics

We present in this work the application of a set Evolutionary Electronics extends the concepts of
of different evolutionary methodologies in the prob- Genetic Algorithms to the evolution of electronic
lem of electronic filter design. The main objectives circuits. The main idea behind Evolutionary Elec-
are to find out which constraints in the filter topolo- tronics is that each possible electronic circuit can be
gies, i f any, must be observed aEong the evolutionary represented as an individual or a chromosome of an
process and t o study the problem of convergence t o evolutionary process, which performs standard ge-
parsimonious circuits. The new area of Evolutionary netic operations over the circuits, such as selection,
Electronics is introduced, an evolutionary methodol- crossover and mutation [13]. We may visualise the
ogy based on variable length representation is pre- task of Evolutionary Electronics as the one of sam-
sented and the results on the evolution of low-pass pling a search space S consisting of electronic circuits
and band-pass filters are described. [3]. The sampled search space S has the following
main features:
1. Introduction
1. It is usually very large and formed by circuits
This work focuses on the application of evolution- of different sizes and topologies;
ary systems [4]in engineering design. Particularly, 2. It can be visualised as formed of two sub-
the application of evolutionary techniques in the area spaces [3]: the sub-space of the compliant circuits,
of electronic design and optimisation gave birth to a Sc, and the sub-space of the non-compliant cir-
new and promising area of research, Evolutionary cuits, Snc.
Electronics[l3][6]. The aim of this area is the cre- The main idea expressed by the last feature is that,
ation of new automation design techniques for elec- given a system specification or target, the search
tronic circuits, based on the Darwinian concepts of space will be usually formed by circuits that satisfy
natural selection, recombination and mutation. We the user specifications, contained in S,, and circuits
present and apply in this work evolutionary method- that do not satisfy the required specifications, con-
ologies based on fixed and variable length genotypes tained in &,. The Artificial Evolutionary System
in the problem of passive analog filter design. should converge to a solution in S,. Nevertheless,
This work is divided into six sections; section 2 engineers are not interested in any solution of S, , but
introduces the basic concepts of Evolutionary Elec- only in those solutions which are optimal or quasi-
tronics; section 3 presents some previous works in the optimal in terms of one or more specific criteria, such
area of passive analog filter design; section 4 presents as the number of circuit components, area, speed
our evolutionary methodologies; section 5 describes or its power consumption. Therefore, the Artificial
a series of experiments made by the authors in elec- Evolutionary System must converge to a design con-
tronic filters design; we compare the results for dif- tained in the restricted subspace of the engineering
ferent filter specifications (low-pass and band-pass) solutions, S e , where Se c S,.This can be viewed in
and also for different evolutionary frameworks, i.e., Figure 1, which shows an example of the mapping
using fixed or variable length representations. Sec- between the fitness landscape and the Evolutionary
tion 6 concludes this work. Electronics search space subsets, as well as an exam-

0-9’803-4869-9198 $ IO.OOO1998 EEE 434

 A Fitness
resentation. Though, they have not addressed the
case of searching for new topologies.

/ Using a different evolutionary methodology, Koza

et. al. have made a large number of works apply-
ing the Genetic Programming (GP) methodology in
electronic circuit synthesis. In[9][10][1], they have
applied Genetic Programming and the Circuit Con-
structing Tree methodologies[9] to the evolution of a
Crossover Filter, evolution of a Brick-Wall Low Pass
Figure 1 - Evolutionary Electronics Genome Space Filter, evolution of an Asymmetric Bandpass Filter
and evolution of an operational amplifier. Both the
filter topologies and the component values have been
ple of the genotype - phenotype mapping in Evolu- evolved and the SPICE simulator has been used to
tionary Electronics; in this figure, A E S,, B E S ,, evaluate the circuits.
and C E Se (shaded area). The individual C, with
the highest fitness, suits as an engineering solution. 4. Evolutionary Methodologies
Due to the broad scope of the area, researchers
have been focusing on different problems, such as Based on these previous works and on the main
placement and routing, evolution of digital circuits features needed for ainalog design automation tools,
based on Boolean gates [l2], evolution of sequential we devised three main issues:
circuits [5][12], evolution of passive and active ana- 1. Evolution of the circuit topologies;
log circuits [9][7] [a], evolution of operational ampli- 2 . Evolution of the components nature and val-
fiers [11][1]and transistor size optimisation. Of great ues;
relevance are the works focusing on intrinsic hard- 3. The need to seek for parsimonious solutions.
ware evolution [13], in which evolution is performed The first issue is examined by comparing two
in the own silicon, exploring, therefore, all the phys- experiments (see next section): one in which the
ical properties of the medium. This particular area topologies sampled by the Evolutionary Algorithm
is called Evolvable Hardware [5]. (EA) are constrained1 to be arranged into parallel
meshes and another in which this constraint is re-
3. Previous Works moved and the EAs are able to make any component
arrangement.
We are particularly addressing the case of evolv- The second issue raised is examined by allowing
ing analog circuits based on resistors, capacitors and the EA to choose the nature and the value of the
inductors. This task has already been tackled from components; however, in order to make this applica-
different evolutionary perspectives by other authors tion as near as possible to the real electronics world,
recently. we restrict resistor, capacitors and inductors to take
J.B. Grimbleby, from the University of Reading only the so-called pre€erred values [7], that are more
[2], has applied Evolutionary Algorithms (EAs) to commonly manufactured.
synthesise filters for particular frequency and time The third issue is of major importance for our
domain specifications. In his approach, the EAs work, since it is directly related to our evolutionary
are used to find the network topology; the compo- methodology. If we are to use genotypes represented
nent values are determined by numerical optimisa- by strings of fixed length to this problem, we have
tion, without evolution. He has used fixed length to guess, possibly using our previous knowledge, the
string structures and integer number chromosome number of components that are necessary to satisfy
represent ation. a particular filter design specification. Furthermore,
In another application, Horrocks and Spittle, penalties will have to be applied in the fitness func-
from the University of Wales [7], have used Genetic tion for large circuits, if parsimony is also an impor-
Algorithms to deal with the problem of preferred val- tant objective. On the other hand, when using GP,
ues electric components selection, particularly resis- care should be taken to control the depth size of the
tors and capacitors, to implement active filters. In trees.
[8], Horrocks and Khalifa extend this study, taking In order to overcome the problems described
into account parasitic effects in resistors, inductors above, we decided to use an EA in which circuits are
and capacitors. They evolved a low pass filter with 9 represented by strings that are allowed to grow along
components encoded in 9 genes through binary rep- the evolutionary process. We call it the Increasing

435
Length Genotypes Approach, which is based on the quency response specification, Taryeti and the fre-
ideas of the Species Adaptation Genetic Algorithm quency response obtained Outputi . The frequency
(SAGA) [4]; all electronic circuits of the initial popu- axis is uniformly sampled over the band of interest,
lation will have a small number of components, which usually ranging from 0 to 10 kHz. The weight wi
increases gradually along the evolutionary process. takes a maximum value for frequency points inside
Therefore, the EA starts sampling subspaces of par- the passband, an intermediate value for frequency
simonious solutions, going toward subspaces of more points inside the stop band and a minimum value
complex solutions as long as design requirements are for other frequency points. We set wi equal to 20, 10
not satisfied. The selection pressure provides a way and 1 for these three cases respectively. Since we are
to stop the growing process when circuits close to the using the roulette well selection method, the value
specification are found, although care must be taken given by equation 1 had to be inverted. Although
with local optima. this fitness equation worked well, setting the values
Therefore, in addition to crossover and mutation, of the weights has been an interactive and time con-
we have also used the Increasing Length Operator, suming process.
which increases, with a low probability [4], the length In order to evaluate the circuits, we have used,
of each genotype by one gene; each new gene created according to the case, a hand written simulator, the
by this operator is randomly initialised. Particularly, SMASH simulator, from Dolphin corporation, and
our crossover operator did not require maximisation the SPICE simulator.
of similarity between genotype segments, which is a
concept used in the SAGA framework. The muta- 5.1 First Test Set
tion operator actuates over integer values by chang-
ing them, with equal probability, to any other value In the first set we constrained the circuit topolo-
within the alphabet used. Tests have been made with gies to be made up of parallel meshes, each one hav-
both fixed and adaptive mutation rates. In the lat- ing two components. Each gene will encode a mesh
ter, the mutation rate increases as the average geno- of the circuit, defining the nature of the two com-
type length grows. ponents and their values (chosen from the preferred
manufactured values). Therefore, each time a geno-
5 . Case Studies type is increased by one (random) gene, one mesh is
added to the circuit. The initial population consists
We have devised three sets of tests to evaluate of circuits of only one or two meshes and may grow
different evolutionary methodologies: until five meshes. Figures 2 and 3 show the evolved
1. Variable length representation EAs, con- circuits for the low-pass and band-pass specifications
straining the topologies of the filters to meshes; respectively and Figures 4 and 5 show their respec-
2. Fixed length representation EAs, making no tive frequency responses (output voltage in decibels
constraints to the filter topologies; versus frequency in Hertz).
3. Variable length representation EAs, making
We have used a population of 50 individuals, run-
no constraints to the filter topologies.
ning along 500 generations, crossover rate of 0.7,
In order to evaluate the above evolutionary
adaptive mutation rate, and increasing length oper-
methodologies, we have used two test cases, a low-
ator rate of 0.05. Fitness proportional selection and
pass and a band-pass filter, with the following fre-
one-point crossover have been used. Each genotype
quency requirements:
is formed of up to 30 positions (5 genes made up of
1. Low-pass filter with passing band up to 1000
6 loci). An integer number representation has been
Hz and stop band above 2000 Hz [lo];
used [2] and, according to this representation, each
2. Band-pass filter with passing band between
gene may represent a total of 82,944 meshes (differ-
2000 Hz and 3000 Hz and stop band below 1000 ent combinations of capacitors, resistors and induc-
Hz and above 4000 Hz. tors of different values). A hand-written circuit sim-
Based on Koza’s previous works [9][10], we have ulator based on the Laplace analysis has been used.
devised the following fitness evaluation function:
5.2 Second Test Set
n
fitness = C wilTarget; - OutputiI (1) In the second test set we have used fixed length
i=l
representation, though allowing any kind of connec-
According to equation 1, the fitness is the tions among the components. Each gene in the geno-
weighted sum of the deviations between the fre- type now describes a single component, by determin-

436
 PI0 0-

Figure 2 - Evolved Low Pass Filter (First Set) Figure 6 - Evolvedl Low Pass Filter (Second Set)

~~

c L
Figure 3 - Evolved Band Pass Filter (First Set) Figure 7 - Evolved Band Pass Filter (Second Set)

ing its nature (resistor, capacitor or inductor), value resistors and inductors).
(from the manufactured preferred values) and con- Figures 6 and 7 show the evolved circuits for the
nections points. We established that all genotypes low-pass and band-pass specifications respectively
would be made up of 10 genes, i. e., each circuit and Figures 8 and 9 show their respective frequency
with 10 components. The SMASH simulator has responses. These circuits have also been simulated
been used to evaluate the circuits, in the AC analysis in SPICE, showing the same results. Note that the
mode. As any kind of arrangement among the com- low-pass filter has only 9 components, because one
ponents is now possible, many topologies are now component was found to have no effect in the circuit
not simulatable, receiving a negative fitness value of behaviour.
-1000 [lo].
5.3 Third Test Set
We have used an Evolutionary Algorithm with 40
individuals, running over 200 generations, crossover In the third test set we have allowed any topol-
rate of 0.7, mutation rate of 0.02 per genotype posi- ogy arrangement between components, using a vari-
tion, linear rank selection with truncation, and uni- able size representation. Each gene represents a par-
form crossover. Each genotype is formed by 50 posi- ticular component, in the way described in the last
tions (10 genes made up of 5 loci). An integer num- section. We initialised all the individuals in the pop-
ber representation has been used and, according to ulation with five random genes and applied an es-
this representation, each gene may represent a total pecial operator to iincrease the genotypes along the
of 14,112 components (different values of capacitors, evolutionary process. Opposing to the first set of ex-

v . D
-O t 1
Figure 4 - Frequency Response for the Circuit of Fig. 2

437
 m>-x
-=-
Figure 10 - Evolved Low Pass Filter (Third Set)
Figure 12 - Frequency Response for the Circuit of Fig. 10

Blob. Lc0.n.w >Sob.

- vm I,&*. <,if

Figure 11 - Evolved Band Pass Filter (Third Set)

periments, in which the unit of increase was a mesh

Figure 13 - Frequency Response for the Circuit of Fig. 11
made up of two components, in this case the unit of
increase was one component. The SMASH simula- 5.4 Summary of the Results
tor has been used to evaluate the circuits in the AC
analysis mode. For the sake of comparison, Tables I and I1
We have used a population of 40 individuals, run- summarise the results obtained over the experiments
ning along 300 generations, crossover rate of 0.7, mu- performed. In these tables, NC is the number of
tation rate of 0.02 per genotype position and increas- components of the best solution and A(lK), A(2K),
ing length operator rate of 0.1. Linear rank selection A(3K) and A(4K) are the response of the circuits,
with truncation, uniform crossover and integer num- in dB, for the frequencies of interest, i. e., 1 kHz,
ber representation have been used. The genotypes 2 kHz, 3 kHz and 4 kHz. For the low pass filter,
are initialised with five genes, each gene being made good designs should have high A(1K) values and low
up of 5 positions and being able to express ti total of A(2K) values; for the band-pass filter, good designs
14,112 components. They are allowed to grow up to should have high A(2K) and A(3K) values and low
a total of 15 genes. A(1K) and A(4K) values. The lines labelled 1, 2
and 3 accounts for the first, second and third set of
Figures 10 and 11 show the evolved circuits for
experiments respectively. It can be seen from these
the low-pass and band-pass specifications respec-
tables that, whereas the three approaches perform
tively and Figures 12 and 13 show their respective
similarly in both cases, the first approach arrived at
frequency responses. These circuits have also been
simulated in SPICE, showing the same results.
Figures 14 and 15 show the average genotype
lengths, in terms of number of electric components,
of the population along the evolutionary process for
particular runs on the low-pass and band-pass cases
respectively. It can be seen from the figures that
there is a tendency for the genotypes to grow along
the evolutionary process, due to the application of
the Increasing Length Operator. Although the In- Figure 14 - Average genotypes length (lowpass filter)
creasing Length operator is applied at a constant
rate along the evolutionary process, we can verify 1PO .
*.a -
from the figures that there are periods in which the =%
z 1-.
d

average length of the genotypes stays stable. In these

periods, the EA is failing to find fitter solutions in
less parsimonious subspaces; so, selection pressure
keeps the average length of the population constant.
This fact suggests that our approach is capable of
finding more parsimonious solutions. Figure 15 - Average genotypes length (bandpass filter)

43 8
 TABLE I
components, in order to make this tool suitable to
SUMMARY FILTER
ON LOW-PASS EVOLUTION
be applied in integrated circuits design.

7. Acknowledgements
The authors would like to thank CAPES, Brazil-
ian Federal Agency, for sponsoring this work.
TABLE I1 IiEFERENCES
FILTER
SUMMARY ON BAND-PASS EVOLUTION Bennett 111,F. H., Koza, J. R., Andre, D., Keane, M. A.,
" Evolution of a 60 Decibel Op Amp Using Genetic Pro-
gramming", Proceedings of the First International Con-
ference on Evolvable Systems (ICES96), LNCS 1259, pp.
455-469, Tsukuba, Japan, October, 1996.
Grimbleby, J. B. , "Automatic Analogue Network Syn-
thesis Using Genetic Algorithms", Proceedings of the
First IEE/IEEE Ihternational Conference on Genetic
Algorithms in EngLineering Systems (GALESIAS - 95),
pp.53-58, UK, 1995.
a smaller circuit in the first case. Harvey, I., Thompson, A., "Through the Labyrinth Evo-
lution Finds a Way: A Silicon Ridge", Proceedings of
Comparing the performance of this evolutionary the First International Conference on Evolvable Systems
methodology with the genetic programming method- (ICES96), LNCS 1259, pp. 406-422, Tsukuba, Japan, Oc-
ology in the evolution of the low-pass filter [lo], tober, 1996.
Harvey, I., "The Artificial Evolution of Adaptive Be-
it has been verified that this methodology arrived haviour", PhD Thesis, University of Sussex, School of
at circuits with similar frequency responses, though Cognitive and Computing Sciences, September, 1993.
processing fewer individuals than in the genetic pro- Higuchi, T., Iba, H., Manderick, B., "Evolvable Hard-
ware", in Massively Parallel Artificial Intelligence (ed.
gramming methodology. H. Kitano), MIT Press, 1994.
In terms of time considerations, the first set of ex- Hirst, T. , "Evolutionary Signal Processing: A Prelimi-
nary Report", Proceedings of the Fourth Conference on
periments are also faster, because we could write a Artificial Life, pp. 425-431, Brighton, UK, July, 1997.
simulator in C to evaluate the circuits, since they are Horrocks, D. H, Spittle, M. C., "Component Value Se-
made up of simple meshes. For the sake of compar- lection For Activate Filters Using Genetic Algorithms",
1st On-line Workshop on Soft Computing (WSCl), Spe-
ison, the written simulator based on Laplace analy- cial Session on Evolutionary Electronics, August 19-30,
sis takes around 1 second to evaluate a circuit, while 1996.
standard simulators used in the second and third sets Horrocks, D. H., Khalifa, Y. M. A., "Genetic Algorithm
Design of Electronic Analogue Circuits Including Par-
take around 3 seconds to do the same. asitic Effects", 1st On-line Workshop on Soft Comput-
Even though the solutions obtained in the three ing (WSCl), Special Session on Evolutionary Electronics,
August 19-30,1996.
sets of experiments are satisfactory in terms of per- Koza J. R., Benne1,t I11 F. H., Andre, D. , Keane, M. A.,
formance, a problem remains to be addressed: the "Four Problems for which a Computer Program Evolved
circuits obtained in all the tests have presented large By Genetic Programming is Competitive with Human
Performance", Proc. Of 1996 IEEE International Confer-
inductor values, which may turn difficult the use ence on Evolutionary Computation, IEEE Press., Pages
of these circuits in real electronic applications. Al- 1-10.
though the first methodology has outperformed the [lo] Koza J. R., Bennett I11 F. H., Andre, D. , Keane, M.
A, "Toward Evolution of Electronic Animals Using Ge-
other ones when time is taken into account, the is- netic Programming", Artificial Life V: Proceedings of the
sue of component size minimisation may change the Fifth International Workshop on the Synthesis and Sim-
situation. When tuning the fitness evaluation func- ulation of Living Systems. Cambridge, MA: The MIT
Press, 1996.
tion to search also for small sized components, it is Ill] Kruiskamp, W. and Leenaerts, D., "DARWIN: CMOS
likely that the new topologies, potentially found in opamp synthesis by means of genetic algorithm", Pro-
the second and third sets, may be more successful. ceedings of the 32nd Design Automation Conference,
New York, NY: Association for Computer Machinery, pp.
698-703,1995.
6. Conclusions [12] Miller, J. F., Thomson, P. , "Combinationaland Sequen-
tial Logic Optimization Using Genetic Algorithms", Pro-
Different evolutionary approaches have been ap- ceedings of the First IEE/IEEE International Conference
on Genetic Algorithms in Engineering Systems ( G A L E
plied to the problem of passive electronic filter de- SIA - 95), pp.34-38, UK, 1995.
sign. The issues of variable and fixed EA representa- [13] Thompson, A., "An Evolved Circuit, Intrinsic in Sili-
con, entwined with physics", Proceedings of the First In-
tion and of imposing constraints in the circuit topolo- ternational Conference on Evolvable Systems (ICES96),
gies have been investigated. We are now studying the Tsukuba, LNCS l259, pp. 390-405, Japan, October,
problem of evolution of topologies with small sized 1996.

439