Discussion Paper No.

02-41

Genetic Algorithms: A Tool for

Optimization in Econometrics
Basic Concept and an
Example for Empirical Applications

Dirk Czarnitzki and Thorsten Doherr

ZEW
Zentrum fr Europische
Wirtschaftsforschung GmbH
Centre for European
Economic Research
 Discussion Paper No. 02-41

Genetic Algorithms: A Tool for

Optimization in Econometrics
Basic Concept and an
Example for Empirical Applications

Dirk Czarnitzki and Thorsten Doherr

Download this ZEW Discussion Paper from our ftp server:

ftp://ftp.zew.de/pub/zew-docs/dp/dp0241.pdf

Die Discussion Papers dienen einer mglichst schnellen Verbreitung von

neueren Forschungsarbeiten des ZEW. Die Beitrge liegen in alleiniger Verantwortung
der Autoren und stellen nicht notwendigerweise die Meinung des ZEW dar.

Discussion Papers are intended to make results of ZEW research promptly available to other
economists in order to encourage discussion and suggestions for revisions. The authors are solely
responsible for the contents which do not necessarily represent the opinion of the ZEW.
Nontechnical Summary
Many new inventions in the field of engineering sciences are based on the
knowledge of structures in nature. These highly efficient structures are
the results of an optimization process called evolution. Evolution is the
strategy used by nature to optimize the adaptation of life according to the
environment.

In this paper, we present a tool for optimization of econometric mod-

els based on evolutionary programming: a genetic algorithm. Genetic
algorithms simulate evolution for a population of candidate solutions
in an artificial environment representing a specific problem. Examples
are the optimization of movement patterns for artificial lifeforms, the
determination of weights for neural networks or, not surprisingly, the
emergence of markets in the economy. These by no means exhaustive
examples demonstrate the versatility of genetic algorithms.

We briefly describe the fundamental concept of genetic algorithms and

explain the design of a specifically developed algorithm which we have
designed to estimate econometric models, especially those models where the
criterion function is not continuously differentiable. This genetic algorithm
is implemented in the statistical software package STATA (Version 7.0).

To demonstrate its performance, we apply it to a difficult econometric prob-

lem: the semiparametric estimation of a censored regression model. We
carry out some Monte Carlo simulations and compare the genetic algo-
rithm with another estimation technique, which is called iterative linear
programming algorithm (ILPA), to run the censored least absolute devia-
tion (CLAD) estimator. It turns out that both algorithms lead to similar
results in this case, but that the proposed method is computationally more
stable than its competitor.
 Genetic algorithms: A tool for optimization in
econometrics Basic concept and an example for
empirical applications1
by

Dirk Czarnitzki and Thorsten Doherr

July 2002

Abstract
This paper discusses a tool for optimization of econometric mod-
els based on genetic algorithms. First, we briefly describe the con-
cept of this optimization technique. Then, we explain the design of a
specifically developed algorithm and apply it to a difficult economet-
ric problem, the semiparametric estimation of a censored regression
model. We carry out some Monte Carlo simulations and compare the
genetic algorithm with another technique, the iterative linear program-
ming algorithm, to run the censored least absolute deviation estimator.
It turns out that both algorithms lead to similar results in this case,
but that the proposed method is computationally more stable than its
competitor.

Keywords: Genetic Algorithm, Semiparametrics, Monte Carlo Simulation

JELClassification: C14, C25, C45, C61, C63

Address: Centre for European Economic Research (ZEW)

Department of Industrial Economics and International Management
P.O.Box 10 34 43
68034 Mannheim
Germany
Phone: +49/621/1235158, 291
Fax: +49/621/1235170
E-Mail: [email protected], [email protected]

1
Helpful comments by Francois Laisney and James L. Powell are gratefully acknowl-
edged. Moreover, we would like to thank the participants of the Econometrics Lunch of
the University of California at Berkeley, where an earlier version of this paper has been
presented.
1 Introduction
Many new inventions in the field of engineering sciences are based on the
knowledge of structures in nature. These highly efficient structures are the
results of an optimization process called evolution. Evolution is the strategy
used by nature to optimize the adaptation of life according to the environ-
ment. The basic principles are crossover, mutation and selection. Evolution
theory explains these principles on the basis of whole populations. Genetic
science takes a much closer look at the individual aspects of evolution: the
genes. It explains the meaning of crossover and mutation at a molecular
level. The knowledge of both worlds is combined in genetic algorithms to use
the problem solving capabilities of evolution for a large number of scientific
and engineering problems or models. Genetic algorithms (GAs) simulate
evolution for a population of candidate solutions in an artificial environment
representing a specific problem. Examples are the optimization of movement
patterns for artificial lifeforms, the determination of weights for neural
networks or, not surprisingly, the emergence of markets in the economy.
These by no means exhaustive examples demonstrate the versatility of GAs.

GAs have already been used in economic research, especially in industrial

economics. For example, Arifovic (1994) applied a GA to the cobweb
model and compares its performance with three other learning algorithms
(least squares learning algorithm, algorithm in which agents form cobweb
expectations as well as sample averages of past prices). She concludes that
the GA converges to the rational expectations equilibrium for a wider range
of parameter values than the competitors. One main result is that the GA
requires less prior information: while other algorithms assume that agents
know how to maximize their objective function, this is not necessary for the
GA. Instead, the agents learn how to maximize their objective functions in
view of their environment. Arifovic points out that this feature of the GA
behaviour is similar to the characteristics of human behaviour observed
in experiments. Price (1997) uses GAs to model strategic behaviour in
several standard industrial organization games. He demonstrates that GA
performs well as a modeling tool and that evolutionary programming has
a potential role in applied economics when detailed market simulations are
required. Cooper (2000) applies a GA to some design problems of firms
research and development activities. He explores stochastic learning curves,
patent design and the importance of technodiversity in the introduction
of new technologies to developing countries. An empirical application has
been conducted by Varetto (1998), who introduces a GA to analyze the

1
insolvency risk of 3,840 industrial Italian companies. Varetto compares
linear discriminant analysis as a traditional statistical methodology for
bankruptcy classification and prediction and a genetic algorithm. He con-
cludes that genetic algorithms are a very effective instrument for insolvency
diagnosis, although the results obtained with linear discriminant analysis
were superior to those obtained with the GA. Varetto notes that the results
of the GA were obtained in less time and with more limited contributions
from the financial analyst than the linear dicriminant analysis.

The use of GAs in econometrics is rare. Especially, for semiparametric esti-

mations it may be a useful alternative to existing optimization algorithms.
The only study we are aware of is Dorsey and Mayer (1995). They apply
GAs to several optimization problems in econometric estimation and carry
out Monte Carlo simulations. Among other methods, Dorsey and Mayer
compare a GA to the Maximum Score estimator developed by Manski
(1975). However, they use data from a study of McManus (1985) on the
effect of capital punishment which consist of only 44 observations. Dorsey
and Mayer (1995) focus their study on the impact of varying the tuning
parameters of a GA but they do not present results on the six coefficients
that are estimated. In line with Dorsey and Mayer, we provide Monte
Carlo results to demonstrate that a GA works well. We estimate a censored
least absolute deviation (CLAD) model using a GA for optimization. We
report results for different sample sizes and analyze the convergence of the
estimates.

The next section explains the concept of GA and introduces basic termi-
nology used in evolutionary programming. Section 3 describes the design
of a genetic algorithm which we have developed for the estimation of
econometric models. In section 4, we apply this GA to a specific econo-
metric model and carry out simulations to study the performance of the GA.

2 Genetic algorithms
The concept of genetic algorithm is very appealingly described by Cooper
(2000). He calls it a concept of partial imitation and refers to an approach
which is familiar to every economist: [...] an effective method for creating
innovative new models is to combine the successful features of two or
more existing models (Cooper, 2000: 403). That is exactly what Nature
does and what is known as evolution. Learning from Nature and finding

2
improved elements of a complex space is incorporated into a formal method
of optimization called genetic algorithm.

2.1 Terminology
The terminology of GAs is mainly borrowed from biology and evolution
theory to underline the analogies. Each term represents the artificial
implementation of biological or evolutionary concepts, though on a much
simpler level. Because there are always different views on the same item,
a multitude of more biological or more evolutionary inspired realizations
exists.

Population denotes a set of specific entities. In biological terms these

entities are organisms of the same species. In the world of GAs the
organisms are from the species candidate solution living in an artificial
environment defined by a given problem. The first generation of candidate
solutions is a sample of possible parameter combinations for the problem.

Fitness is a term from evolution theory. It is a measure of the sur-

vival and the reproduction probability and fertility of an entity. The
fitness function of the GA defines the environment for the artificial
evolution. Each candidate solution in the population will be evaluated by
the fitness function. A higher fitness results in a higher survival chance
and a higher reproduction rate. A GA maximizes the fitness of a population.

Selection is the evolutionary term for survival of the fittest, referring

to the probability for an organism to survive and reproduce. Most GAs
described in the literature have been generational at each generation
the new population consists entirely of offsprings formed by parents in the
previous generation (Mitchell 1996). The parent generation is completely
discarded. These GAs rely only on selection for reproduction. Other GAs
additionally implement the struggle for survival. The environment can
only support a given number of entities. An evolutionary step of a GA
consists of reproduction to create an intermediate population of parents
and offspring, and of evaluation and selection of the fittest entities to re
establish the original population size. This and similar selection methods
are based on a concept called elitism, first introduced by De Jong (1975).
Many researchers have found that elitism significantly improves the GAs
performance (Mitchell 1996).

3
Reproduction is another basic principle of evolution. In GAs the fertility
of a candidate solution is determined by the relative fitness. A fitter
solution will reproduce more often than a less fit entity.

Deoxyribonucleic Acid (DNA) is the carrier of genetic information for

all complex organisms. The DNA is organized in strings which are called
chromosomes that serve as a blueprint for the organism. Specific sections
of chromosomes define the genes: functional blocks of DNA, each of which
encodes a particular protein. On a more abstract level which we call the
evolutionary view from now on the genes encode different traits, such
as the color of the skin. The most biologically inspired form of artificial
DNA encoding for GAs is the use of segmented bit strings. Each segment
represents a parameter of the fitness function. Mutation and crossover
can easily be achieved by simple binary operations. The drawback of this
method is that a small change in the bit pattern can double or halve the
real value of the segment. Although there have been many extensions
to the basic binary encoding schema, such as gray coding (Bethke, 1980,
Caruna and Schaffer, 1988) to circumvent this problem, binary encoding
is considered unnatural and unwieldy for many problems (Mitchell, 1996).
A more evolution inspired method is the use of real valued encodings.
These are much more natural to use because the fitness functions of many
problems require real numbers as parameter inputs. A candidate solution
is more a set of traits than a chromosome.

Crossover is the recombination of the subsequences from two chromosomes

to create two offsprings. The new chromosomes share the genes of both
parents. From the evolutionary point of view, crossover secures the
continuation of successful traits. In GAs the method to simulate crossover
depends on the method of DNA encoding. Binary encoding requires only
the splitting and recombination of the parent bit streams at randomly
chosen crossover points simulating the biological process. Real value
encoded GAs use more complex methods to simulate the crossover of traits.
Since traits are represented by real numbers the trait of the offspring can
be calculated as a randomly weighted arithmetic mean of the corresponding
trait numbers of the parents. This allows the child to have the hair
of its mother, the eyes of its father and a nose with a little bit of both parents.

4
Mutation is a very important evolutionary aspect for GAs. While crossover
can produce many new variants of existing solutions, mutation has the power
to produce completely new solutions. It is randomly applied after crossover
to mutate one or more genes in an offspring. GAs using binary encoding just
have to apply the logical not operator at randomly chosen bit positions in
the string. The mutation of randomly selected real value encoded traits is
resolved by the multiplication with a random factor within a specific interval
termed for further reference radiation level.

2.2 Implementation of a genetic algorithm

There is no golden rule to implement a GA. The biological and evolutionary
concepts of GA leave much room for interpretation. Many additional
concepts can be introduced to optimize the GA for a specific problem. We
choose a very straightforward approach using real valued encoding and
elitism rooted more in the evolutionary than in the biological soil. The
target of this implementation was to devise a reliable and versatile tool
for many statistical optimization problems. The GA has been developed
in STATA Version 7.0, a statistical software package with a flexible
programming language. The fitness function can be dynamically linked
to the GA, so that there is no need to reprogram the GA for a specific
problem. The integration of the GA into other STATA programs can easily
be achieved. The implementation is described below:

1. Creation of the initial population

The initial population consists of s survivor vectors representing the
candidate solutions. A vector has k elements corresponding to the
parameters of the fitness function

fi = f (i1 , . . . , ik ) with i = 1, . . . , s. (1)

The elements ij (j = 1, . . . , k) are initialized by a random value in a

particular interval [aj , bj ] which has to be chosen by the user. A first
evaluation of the fitness f of each vector is then performed.

2. Main loop
The main loop runs the artificial evolution. It repeats steps 3 to 5 until
a maximum number of generations T is reached or the GA stagnates.
Stagnation occurs when the current generation equals the previous

5
 generation over a given number of subsequent generations (with
T) .

3. Determination of the mutation probability and radiation level

The values of the mutation probability and the radiation level both
shrink according to the general halflife formula:
!
ln(2)
t = 0 exp t with t = 1, . . . , T, (2)

ln(2)
 
t = 0 exp t with t = 1, . . . , T, (3)

where 0 and 0 are the initial values and and are the halflife
durations. Since mutation is a probability, the initial value must be
in the interval [0, 1]. The absolute value of the radiation level and its
negative counterpart define the interval limits for the random mutation
factor.

4. Determination of the selection probability

A selection probability i has to be associated for each vector i . The
first step is the determination of the minimum and the maximum of
the fitness values to calculate an offset for the following normalization
scheme:

max(f1 , . . . , fs ) min(f1 , . . . , fs )
offset = (4)
s

The offset is required to give the lowest fitness a reasonable probability

greater than zero. Therefore, we compute a rescaled fitness

hi = fi min(f1 , . . . , fs ) + offset. (5)

Then, the selection probability is defined as

hi
i = s . (6)
X
hi
i=1

In rare occasions, it may be possible that users want to decrease the

importance of the current fitness for selection, for example, as an addi-
tional tool to minimize the danger of getting trapped in local extrema.
Therefore, we include an optional weight W for the selection probabil-
ity. By default, W = 1 which means that the selection probability for

6
 crossovers is fully determined by the fitness of the candidate solutions.
If W is set to values smaller than 1, the importance of the individual
fitness decreases. If W = 0, the selection probability is independent
of the fitness, so that the chance of being chosen for crossover would
be equal for every candidate solution. If W is specified, the selection
probability is calculated as a the convex combination

1
i = (1 W ) + W i . (7)
s

5. Evolution

Two different candidate solutions are drawn out of the population

according to the selection probabilities i (optionally weighted by
W ).
Crossover is applied by a randomly weighted mean for each el-
ement pair of the drawn candidates. The weight is uniformly
distributed on the interval [0; 1].
An element of the resulting offspring vector is mutated according
to the mutation probability .
The radiation level determines the interval [; ] of the random
mutation factor which is uniformly distributed.
The offspring is evaluated by the fitness function.
This process is repeated until the number of offsprings reaches a
given number o. This intermediate population is sorted by the
fitness of its members to determine the s survivors for the next
generation.
The criterion of stagnation (see step 2 above) is fulfilled when the
new generation contains no offspring.

3 An example of implementing genetic algorithms

in econometrics
This section presents a small MonteCarlo study of the estimation of a
semiparametric econometric model: the censored least absolute deviation
(CLAD) model proposed by Powell (1984). We compare the genetic
algorithm with an estimation technique called iterative linear programming
algorithm (ILPA) suggested by Buchinsky (1994). Note that there is no
particular reason for choosing this model for demonstration of the GA. The

7
GA can be applied to every numerical criterion function. However, as we
encountered problems using the CLAD estimator in empirical studies, we
thought it might be useful to think about the GA as another option for
practical use or at least as a supplementary method.

Consider the econometric model

yi = 0 xi + i , (8)

where xi is a set of regressors, the corresponding coefficient vector to be

estimated and i a stochastic error term. yi is an unobserved variable and
we only observe a left censored variable
(
yi if yi > 0,
yi = (9)
0 if yi 0.

A special case is the Tobit model which is a fully parametric model and
can be estimated with the common maximum likelihood (ML) techniques.
It is derived from the additional assumption that yi N (, 2 ). If the as-
sumptions of homoscedasticity or normality are violated, the ML estimates
may be inconsistent. In case of heteroscedasticity, researchers can attempt
to model heteroscedasticity as a function of some observable variables.
However, the true functional form is usually unknown and the choice
of variables determining the heteroscedasticity function is arbitrary. To-
bit estimates are sensitive to different choices of the heteroscedasticity term.

Powell has developed different semiparametric models to relax the strict

assumptions which were needed to estimate censored regression models.
Among others, he proposed the CLAD model, where he suggests to esti-
mate M ED(yi |xi ) instead of its expectation. This quantile regression can
be expressed by the minimization problem
N 
yi max(0, 0 xi ) .
X 
CLAD = argmin (10)
i=1

This estimator is consistent and asymptotically normally distributed even

in case of heteroscedastic and/or nonnormally distributed error terms (see
Powell 1984, 1994). However, as this criterion function is not continuously
differentiable, estimation is not easy. Buchinsky (1994) has proposed an
iterative technique (ILPA) which uses the idea of sample trimming to es-
timate CLAD models. His procedure can be summarized by the following
steps:

8
 1. Estimate a median regression (least absolute deviation: LAD) with
the entire sample and generate the estimated yi for this initial step.

2. Subsequently, the sample is trimmed, i.e. the observations for which

yi < 0 are dropped. It is noteworthy that we keep observations for
which yi 0. In Buchinskys original procedure, he suggested to
keep only those observations for which yi > 0. However, Fitzenberger
(1994) has shown that the modified version (using the subset of obser-
vations if yi 0) is more likely to converge than Buchinskys original
proposal. Fitzenberger calls it modified iterative linear programming
algorithm (MILPA).

3. Repeat the estimation of a median regression for the trimmed sample

and predict yi again (for the entire sample).

4. Return to step 2 and keep iterating until the estimated coefficients do

not change during the iterations.

Buchinskys algorithm has the advantage that it is quite easy to implement

using standard econometric software packages (which provide median
regressions as implemented command). Fitzenberger (1994) shows that
ILPA is less likely to converge than his modification MILPA. Moreover,
he finds that both ILPA and MILPA are outperformed by another algorithm
called BRCENS: Fitzenberger adapts an algorithm (BarrodaleRoberts) for
standard LAD regressions and extends it to the CLAD model. It outper-
forms the others with respect to the frequencies that the global optimum is
reached. However, conditional on convergence of ILPA and BRCENS there
is no clear ranking between the algorithms (see also Fitzenberger 1997 and
Fitzenberger and Winker 1999).

The practical problem of nonconvergence of the Buchinsky algorithm which

occurs in applications is that the iteration procedure may jump between two
(or more) trimmed states of the sample. If it is running in circles it is unclear,
what the researcher can do to either achieve convergence or how to decide
whether one of these circling states represents a minimum. Of course, one
could record the value of the criterion function and choose the minimum
but there is no reason which ensures that this value is a global optimum. It
is noteworthy that this circling of the algorithm is not unlikely to happen.
Due to these phenomenon, we compare the performance of the (modified)
Buchinsky algorithm called MILPA with the genetic algorithm in censored
regression models. First, we choose a very simple case of a model and carry

9
out MonteCarlo simulations. Consider the true model
yi = 1 x1i + 2 x2i + 3 x3i , i = 1, . . . , N, (11)
with the constant term x3i = 1 i. The explanatory variable x1 is stan-
dard normally distributed and x2 is uniformly distributed on the interval
[0,1]. For simplicity, we set 1 = 2 = 3 = 1. The observed dependent
variable is
(
yi + i if yi + i > 0
yi = (12)
0 if yi + i 0
where i is a normally distributed heteroscedastic error term. The het-
eroscedasticity is modeled as
i = exp(0.2x4 ), with x4 N (0, 1). (13)
Table 1 displays the required tuning parameters for running the GA to
estimate 1 , 2 and 3 . The tuning parameters have to be set by the user.
The right column shows their values for the following Monte Carlo study.
The initialization interval for all three parameters was set to [2; 2].

Table 1:
Set of tuning parameters for running the genetic algorithm
Parameter Description Parameter value
s Population size 30
o Number of offsprings 60
0 Mutation probability 0.5
0 Radiation level 1
Halflife duration of mutation 15
Halflife duration of radiation 15
T Maximum number of generations 250
Number of subsequent stagnations 10
W Weight for the selection probability 1.0
Initially, we start with a sample of 400 observations and carry out 200
replications of each regression (Buchinsky and GA) to obtain distributions
of the estimated coefficients. Afterwards, we repeat the simulation with
samples of 800, 1,600 and 3,200 observations to investigate the convergence
behavior of the algorithms.

The results suggest that both the Buchinsky algorithm and the genetic
algorithm produce consistent estimates, although the findings are not
exactly numerically identical. Table 2 displays the mean squared errors of
both methods for all simulations.

10
 Table 2:
Mean squared errors of estimated coefficients (multiplied with 100)
Regressor number of observations in sample
400 800 1,600 3,200
x1 GA 0.4244 0.2885 0.1468 0.0686
Buchinsky 0.4207 0.2870 0.1468 0.0689
x2 GA 5.8185 2.7720 1.2521 0.6561
Buchinsky 5.9725 2.2818 1.3054 0.6497
x3 GA 1.7721 1.0356 0.4634 0.2209
Buchinsky 1.8326 0.9096 0.4833 0.2158

However, it is noteworthy that the Buchinsky algorithm fails to converge

in about 12% of all estimations. In such cases, we have dropped 1% of the
observations of the initial sample and repeated the estimation. Usually, the
Buchinsky algorithm converges after dropping a few observations. However,
in empirical applications with real data, this may cause problems as
information is lost.

Figures 1 to 3 display kernel density estimates of the distributions of 1 , 2

and 3 . Testing for normality of the estimates shows that the hypothesis
both algorithms produce asymptotically normal distributed results has
not to be rejected.2 Figure 4 shows the same results as presented in Figures
1 to 3, but displayed as box plots of the coefficients distributions. The
line in the middle of each box represents the median of the data. The box
extends from the 25th percentile to the 75th percentile, the interquartile
range (IQR). Additionally, STATA computes whiskers which extend to the
upper and lower adjacent values. The upper adjacent value is defined as
the largest data point less than or equal to 75th percentile + 1.5 IQR.
The lower adjacent value is defined as the smallest data point greater than
or equal to 25th percentile 1.5 IQR. Observed points more extreme
than the adjacent values, if any, are referred to as outside values and are
individually plotted. (STATA Graphics Manual, Release 6, p. 32).

2
Test results are not presented here, but are available upon request.

11
Figure 1: Gauss kernel density estimates of 1

12
 Figure 2: Gauss kernel density estimates of 2

We have dropped one data point from the results of the genetic
algorithm. For the sample with 800 observations one estimated
value was -.0027. As this would influence the graphical presen-
tation of all other results, we have decided to exclude it in this
figure.

13
Figure 3: Gauss kernel density estimates of 3

14
Figure 4: Box plots of estimated coefficients distributions
Coefficient of x1

Coefficient of x2

Coefficient of x3

15
As the figures above show, both algorithms produce very similar distribu-
tions of results. With growing sample size, the estimates seem to converge
to the true parameter values. From econometric theory, it is known that the
1
CLAD model should exhibit a rate of convergence of N 2 (Powell, 1984).
However, for the genetic algorithm this behavior is unclear a priori as no
theoretical econometric foundation exists. To investigate the convergence
behavior in more detail, we calculate the estimates variances N,k and
relate them to the sample sizes. Searching for k such that

N,k
= constant k = 1, . . . , 3, N = 400, 800, 1600, 3200,
N k
we can estimate the rate of convergence with the following regression model

ln(
N,k ) = k ln(N ) + constant + k .

For all three k the rate of convergence the estimate should be about
0.5. To analyze the rate of convergence of the model empirically, we com-
pare the findings of both methods. The regressions yield the following coef-
ficients:

Table 3:
Empirical rate of convergence of both methods
convergence k for Genetic algorithm Buchinsky
1 .4466 (.0442) .4425 (.0454)
2 .5286 (.0115) .5210 (.0331)
3 .5068 (.0257) .5065 (.0147)
Standard errors in parentheses.

For all six cases, tests show that the hypothesis that the slope coefficient is
0.5 does not have to be rejected.3 Hence, we conclude that both algorithms
exhibit the same behavior of convergence.

4 Some practical experience

In this section we discuss some of our experience in using the GA for econo-
metric applications. First, the tuning parameters are an important feature
of the GA which may influence its performance.

1. Population size and number of offsprings

In the beginning of our study, we used to create a large population, e.g.
3
The reliability of these results may be questioned because we have only four obser-
vations from our simulations for each regression and we estimate two parameters, the
constant term and the slope coefficient which represents the rate of convergence k .

16
 of 100 candidate solutions and chose a smaller number of offsprings,
e.g. 50. This makes the process of evolution sluggish as only a few
crossovers take place, which means that it may take several generations
until the fitness of the best candidate solution improves significantly.
We recommend using a smaller population size, e.g. 30, and to create
a larger number of crossovers, e.g. 60, as done in the presented simula-
tion. This may require more generations until the evolution stagnates,
but it is rapidly computed and the best fitness is found quickly due to
a large number of trials (offsprings) to find improvements.

2. Mutation and radiation

The mutation probability is an important feature to avoid local ex-
trema. So it should not be set to zero or close to zero. Although
we do not provide simulations on the effect of changing the radiation
level and the halflife time of radiation and mutation, we have exper-
imented with these parameters. The halflife time had no perceptible
influence on the performance of the GA. However, we recommend us-
ing a reasonable value for each parameter, because this reduces the
danger of getting stuck in local extrema. A higher level of radiation
yields many trials which may be further away from possible good so-
lutions and thus increases the number of necessary generations until
stagnation.

3. Initial parameter interval

The initial parameter intervals are needed to create the first popu-
lation. Although we found that the mutation feature is a powerful
option to get out of wrong starting intervals, we recommend to set
the initial intervals to meaningful positions and widths. If the true
solutions are outside the initial interval the GA will usually still find
the global optima due to the mutation feature, but the convergence is
slower.

4. Weight of the selection probability

The weight of the selection probability W should be a rarely used fea-
ture. By default, W is set to 1. In our case, there was no need to scale
down the importance of the current fitness for the selection probabil-
ity. We did not encounter problems of ending up in local extrema.
Usually, the evolutionary process finds the global solution. However,
there may be some special cases, where the selection is desired to be
completely random. For example, if multiple solutions are expected.

17
5 Conclusions and Perspectives
In this paper, we present the concept of a genetic algorithm. This method
of optimization is based on Nature and simulates the evolution of artificial
lifeforms. We propose to use genetic algorithms as an alternative tool for
optimizing criterion functions, especially those which are not continuously
differentiable because many alternative techniques run into numerical
problems in this case.

We present the implementation of a genetic algorithm developed by us and

programmed in STATA (Version 7.0). As an example of its performance,
we do some Monte Carlo simulations on the semiparametric censored least
absolute deviations model (CLAD). We compare the results of the genetic
algorithm with those of the algorithm suggested by Buchinsky (1994). We
conclude that both results are similar: the mean squared errors are of equal
size, the distributions of estimated coefficients look similar and the rate of
convergence is almost identical in our simulations. However, the genetic
algorithm is numerically more stable than its competitor.

For future applications, it would be interesting to compare the genetic al-

gorithm with other estimation methods: for example, a simulation on the
difficult optimization of maximum score models (Manski 1975, 1985, Man-
ski and Thompson 1986) including the comparison between the genetic al-
gorithm and the original technique for optimization suggested by Manski.
Finally, an application with nonsimulated data would be desirable.

Notice for interested researchers

The genetic algorithm is programmed in STATA Version 7.0 as an adofile.
If you are interested in using it yourself, the genetic algorithm is available
for download at
ftp://ftp.zew.de/pub/zew-docs/div/genetic.zip
If you encounter problems with the download, please do not hesitate to
contact us by e-mail, either [email protected] or [email protected]. We will
send the adofile and a simple example for running the genetic algorithm.

18
References
Arifovic, J. (1994), Genetic Algorithm learning and the cobweb model,
Journal of Economic Dynamics and Control 18, 328.

Bethke, A.D. (1980), Genetic algorithms as function optimizers, Ph.D. the-

sis, University of Michigan, Ann Arbor.

Buchinsky, M. (1994), Changes in the U.S. wage structure 19631987: Ap-

plication of quantile regression, Econometrica 62(2), 405458.

Caruana R.A. and J.D. Schaffer (1988), Representation and hidden bias:
gray versus binary coding for genetic algorithms, Proceedings of the
fourth international conference on machine learning.

Cooper, B. (2000), Modelling research and development: How do firms

solve design problems?, Journal of Evolutionary Economics 10, 395
413.

De Jong, K.A. (1975), An analysis of the behavior of a class of genetic

adaptive systems, Ph.D. thesis, University of Michigan, Ann Arbor.

Dorsey, R.E. and W.J. Mayer (1995), Genetic algorithms for estimations
problems with multiple optima, nondifferentiability, and other irregu-
lar features, Journal of Business & Economic Statistics 13(1), 5366.

Fitzenberger, B. (1994), A note on estimating censored quantile regressions,

Discussion Paper 14, University of Konstanz.

Fitzenberger, B. (1997), A guide to censored quantile regressions, in: G.

Maddala and C. Rao (eds.), Handbook of statistics, Vol. 15: robust
inference, Amsterdam, 405437.

Fitzenberger, B. and P. Winker (1999), Improving the Computation of

censored quantile regressions, Discussion Paper 56899, University of
Mannheim.

Manski, C.F. (1975), Maximum score estimation of the stochastic utility

model of choice Journal of Econometrics 3, 205228.

Manski, C.F. (1985), Semiparametric analysis of discrete response: Asymp-

totic properties of the maximum score estimator, Journal of Econo-
metrics 27, 313333.

19
Manski, C.F. and T.S. Thompson (1986), Operational characteristics of
maximum score estimation, Journal of Econometrics 32, 85108.

McManus, W.S. (1985), Estimates of the deterrent effect of capital punish-

ment: the importance of researchers prior beliefs, Journal of Political
Economy 93, 417425.

Mitchell, M. (1996), An introduction to genetic algorithms, Cambridge.

Powell, J.L. (1984), Least absolute deviations for the censored regression
model, Journal of Econometrics 25, 303325.

Powell, J.L. (1994), Estimation of semiparametric models, in: R.F. Engle

and D.L. McFadden, Handbook of econometrics IV, New York.

Price, T.C. (1997), Using coevolutionary programming to simulate strate-

gic behaviour in markets, Journal of Evolutionary Economics 7, 219
254.

STATA Corporation (1999), STATA Graphics Manual, Release 6, College

Station.

Varetto, F. (1998), Genetic algorithms applications in the analysis of in-

solvency risk, Journal of Banking and Finance 22, 14211439.

20