geatbx.com/docu/algindex.

html

Evolutionary Algorithms 1 Introduction

1 Introduction
Fig. 1-1: Problem solution using evolutionary algorithms

Different main schools of evolutionary algorithms have evolved during the last 40 years:
genetic algorithms, mainly developed in the USA by J. H. Holland [Hol75], evolutionary
strategies, developed in Germany by I. Rechenberg [Rec73] and H.-P. Schwefel [Sch81]
and evolutionary programming [FOW66]. Each of these constitutes a different approach,
however, they are inspired by the same principles of natural evolution. A good introductory
survey can be found in [Fdb94a].

This document describes algorithms of evolutionary algorithms. In Chapter 2 a short

overview of the structure and basic algorithms of evolutionary algorithms is given.
Chapter 3 describes selection. In Chapter 4 the different recombination algorithms are
presented. Chapter 5 explains mutation and Chapter 6 reinsertion.

Chapter 8 covers parallel implementations of evolutionary algorithms especially the regional

population model employing migration in detail. The application of multiple/different
strategies during an optimization including competition between subpopulations is
discussed in Chapter 9.

Chapter 10 explains how complete optimization algorithms can be created from the
different evolutionary operators. The respective options are discussed in detail. Each of the
presented optimization algorithms represents an evolutionary algorithm.

Chapter 11 lists all the used references and a large number of other publications from the
field of Evolutionary Algorithms.

GEATbx: Genetic and Evolutionary Algorithm Toolbox for use with Matlab - www.geatbx.com.
The Genetic and Evolutionary Algorithm Toolbox is not public domain.
© 1994-2006 Hartmut Pohlheim, All Rights Reserved, ([email protected]).

1/2
2/2
 geatbx.com/docu/algindex-01.html

Evolutionary Algorithms 2 Overview

2 Overview
Evolutionary algorithms are stochastic search methods that mimic the metaphor of natural
biological evolution. Evolutionary algorithms operate on a population of potential solutions
applying the principle of survival of the fittest to produce better and better approximations to
a solution. At each generation, a new set of approximations is created by the process of
selecting individuals according to their level of fitness in the problem domain an d breeding
them together using operators borrowed from natural genetics. This process leads to the
evolution of populations of individuals that are better suited to their environment than the
individuals that they were created from, just as in natural adaptation.

Evolutionary algorithms model natural processes, such as selection, recombination,

mutation, migration, locality and neighborhood. Figure shows the structure of a simple
evolutionary algorithm. Evolutionary algorithms work on populations of individuals instead
of single solutions. In this way the search is performed in a parallel manner.

Fig. 2-1: Structure of a single population evolutionary algorithm

At the beginning of the computation a number of individuals (the population) are randomly
initialized. The objective function is then evaluated for these individuals. The first/initial
generation is produced.

If the optimization criteria are not met the creation of a new generation starts. Individuals
are selected according to their fitness for the production of offspring. Parents are
recombined to produce offspring. All offspring will be mutated with a certain probability. The
fitness of the offspring is then computed. The offspring are inserted into the population
replacing the parents, producing a new generation. This cycle is performed until the
optimization criteria are reached.
1/5
Such a single population evolutionary algorithm is powerful and performs well on a wide
variety of problems. However, better results can be obtained by introducing multiple
subpopulations. Every subpopulation evolves over a few generations isolated (like the
single population evolutionary algorithm) before one or more individuals are exchanged
between the subpopulation. The multi-population evolutionary algorithm models the
evolution of a species in a way more similar to nature than the single population
evolutionary algorithm. Figure shows the structure of such an extended multi-population
evolutionary algorithm.

Fig. 2-2: Structure of an extended multipopulation evolutionary algorithm

From the above discussion, it can be seen that evolutionary algorithms differ substantially
from more traditional search and optimization methods. The most significant differences
are:

Evolutionary algorithms search a population of points in parallel, not just a single

point.
Evolutionary algorithms do not require derivative information or other auxiliary
knowledge; only the objective function and corresponding fitness levels influence the
directions of search.
Evolutionary algorithms use probabilistic transition rules, not deterministic ones.
Evolutionary algorithms are generally more straightforward to apply, because no
restrictions for the definition of the objective function exist.
Evolutionary algorithms can provide a number of potential solutions to a given
problem. The final choice is left to the user. (Thus, in cases where the particular
problem does not have one individual solution, for example a family of pareto-optimal
solutions, as in the case of multi-objective optimization and scheduling problems,
then the evolutionary algorithm is potentially useful for identifying these alternative
solutions simultaneously.)

2/5
The following sections list some methods and operators of the main parts of Evolutionary
Algorithms. A thorough explanation of the operators will be given in the following chapters.

2.1 Selection

Selection determines, which individuals are chosen for mating (recombination) and how
many offspring each selected individual produces. The first step is fitness assignment by:

proportional fitness assignment or

rank-based fitness assignment, see Section 3.1,
multi-objective ranking, see Section 3.23.2.

The actual selection is performed in the next step. Parents are selected according to their
fitness by means of one of the following algorithms:

roulette-wheel selection, see Section 3.3,

stochastic universal sampling, see Section 3.4,
local selection, see Section 3.5,
truncation selection, see Section 3.6 or
tournament selection, see Section 3.7.

For more info see Chapter 3.

2.2 Recombination

Recombination produces new individuals in combining the information contained in the

parents (parents - mating population). Depending on the representation of the variables of
the individuals the following algorithms can be applied:

All presentation:
discrete recombination, see Subsection 4.1, (known from recombination of real
valued variables), corresponds with uniform crossover, see Subsection 4.3.2
(known from recombination of binary valued variables),
Real valued recombination, see Section 4.2:
intermediate recombination, see Subsection 4.2.1,
line recombination, see Subsection 4.2.2,
extended line recombination, see Subsection 4.2.3.
Binary valued recombination, see Section 4.3:
single-point / double-point /multi-point crossover, see Subsection 4.3.1,
uniform crossover, see Subsection 4.3.2,
shuffle crossover, see Subsection 4.3.3,
crossover with reduced surrogate, see Subsection 4.3.4.

For the recombination of binary valued variables the name 'crossover' is established. This
has mainly historical reasons. Genetic algorithms mostly used binary variables and the
name 'crossover'. Both notions (recombination and crossover) are equivalent in the area of
3/5
Evolutionary Algorithms. For consistency, throughout this study the notion 'recombination'
will be used (except when referring to specially named methods or operators).

For more info see Chapter 4.

2.3 Mutation

After recombination every offspring undergoes mutation. Offspring variables are mutated
by small perturbations (size of the mutation step), with low probability. The representation
of the variables determines the used algorithm. Two operators are explained:

mutation operator for real valued variables, see Section 5.1,

mutation for binary valued variables, see Section 5.2.

For more info see Chapter 5.

2.4 Reinsertion

After producing offspring they must be inserted into the population. This is especially
important, if less offspring are produced than the size of the original population. Another
case is, when not all offspring are to be used at each generation or if more offspring are
generated than needed. By a reinsertion scheme is determined which individuals should be
inserted into the new population and which individuals of the population will be replaced by
offspring.

The used selection algorithm determines the reinsertion scheme:

global reinsertion for all population based selection algorithm (roulette-wheel

selection, stochastic universal sampling, truncation selection),
local reinsertion for local selection.

For more info see Chapter 6.

2.5 Population models - parallel implementation

of evolutionary algorithms

The extended management of populations (population models) allows the definition of

extensions of Evolutionary Algorithms. These extensions can contribute to an increased
performance of Evolutionary Algorithms.

The following extensions can be distinguished:

global model, see Section 8.1,

local model (diffusion model, neighborhood model, fine grained model), see
4/5
 Section 8.2,
regional model (migration model, island model, coarse grained model), see
Section 8.3.

For more info see Chapter 8.

2.6 Application of multiple/different strategies

and competition between subpopulations

Based on the regional population model the application of multiple different strategies at the
same time is possible. This is done by applying different operators and parameters for each
subpopulation. For an efficient distribution of resources during an optimization competing
subpopulations are used.

application of multiple strategies, see Section 9.3,

competition between subpopulations, see Section 9.2.

These extensions of the regional population model contribute to an increased performance

of Evolutionary Algorithms, especially for large and complex real-world applications.

For more info see Chapter 9.

This document is part of version 3.8 of the GEATbx: Genetic and Evolutionary Algorithm Toolbox for use
with Matlab - www.geatbx.com.
The Genetic and Evolutionary Algorithm Toolbox is not public domain.
© 1994-2006 Hartmut Pohlheim, All Rights Reserved, ([email protected]).

5/5
 geatbx.com/docu/algindex-02.html

Evolutionary Algorithms 3 Selection

3 Selection
In selection the offspring producing individuals are chosen. The first step is fitness
assignment. Each individual in the selection pool receives a reproduction probability
depending on the own objective value and the objective value of all other individuals in the
selection pool. This fitness is used for the actual selection step afterwards.

Throughout this section some terms are used for comparing the different selection
schemes. The definition of these terms follow [Bak87] and [BT95].

selective pressure

probability of the best individual being selected compared to the average probability
of selection of all individuals

bias:

absolute difference between an individual's normalized fitness and its expected

probability of reproduction

spread

range of possible values for the number of offspring of an individual

loss of diversity

proportion of individuals of a population that is not selected during the selection

phase

selection intensity

expected average fitness value of the population after applying a selection method to
the normalized Gaussian distribution

selection variance

expected variance of the fitness distribution of the population after applying a

selection method to the normalized Gaussian distribution

3.1 Rank-based fitness assignment

1/13
In rank-based fitness assignment, the population is sorted according to the objective
values. The fitness assigned to each individual depends only on its position in the
individuals rank and not on the actual objective value.

Rank-based fitness assignment overcomes the scaling problems of the proportional fitness
assignment. (Stagnation in the case where the selective pressure is too small or premature
convergence where selection has caused the search to narrow down too quickly.) The
reproductive range is limited, so that no individuals generate an excessive number of
offspring. Ranking introduces a uniform scaling across the population and provides a simple
and effective way of controlling selective pressure.

Rank-based fitness assignment behaves in a more robust manner than proportional fitness
assignment and, thus, is the method of choice. [BH91], [Why89]

3.1.1 Linear ranking

Consider Nind the number of individuals in the population, Pos the position of an individual
in this population (least fit individual has Pos=1, the fittest individual Pos=Nind) and SP the
selective pressure. The fitness value for an individual is calculated as:

Linear ranking:

(3-1)

Linear ranking allows values of selective

pressure in [1.0, 2.0].

3.1.2 Non-linear ranking

A new method for ranking using a non-linear distribution was introduced in [Poh95]. The
use of non-linear ranking permits higher selective pressures than the linear ranking method.

Non-linear ranking:

(3-2)

X is computed as the root of the polynomial:

(3-3)

Non-linear ranking allows values

of selective pressure in [1, Nind -
2].

3.1.3 Comparison of linear and non-linear ranking

Figure compares linear and non-linear ranking graphically.

Fig. 3-1: Fitness assignment for linear and non-linear ranking

2/13
The probability of each individual being selected for mating depends on its fitness
normalized by the total fitness of the population.

Table contains the fitness values of the individuals for various values of the selective
pressure assuming a population of 11 individuals and a minimization problem.

Table 3-1: Dependency of fitness value from selective pressure

fitness value (parameter: selective pressure)

linear ranking no ranking non-linear ranking

objective value 2.0 1.1 1.0 3.0 2.0

1 2.0 1.10 1,0 3.00 2.00

3 1.8 1.08 1,0 2.21 1.69

4 1.6 1.06 1,0 1.62 1.43

7 1.4 1.04 1,0 1.99 1.21

8 1.2 1.02 1,0 0.88 1.03

9 1.0 1.00 1,0 0.65 0.87

10 0.8 0.98 1,0 0.48 0.74

15 0.6 0.96 1,0 0.35 0.62

20 0.4 0.94 1,0 0.26 0.53

30 0.2 0.92 1,0 0.19 0.45

3/13
 95 0.0 0.90 1,0 0.14 0.38

3.1.4 Analysis of linear ranking

In [BT95] an analysis of linear ranking selection can be found.

Fig. 3-2: Properties of linear ranking

Selection intensity

(3-4)

Loss of diversity

(3-5)

Selection variance

(3-6)

3.2 Multi-
objective
Ranking

Where proportional and rank-based fitness assignment is concerned it is assumed that

individuals display only one objective function value. In many real world problems,
however, there are several criteria which have to be considered in order to evaluate the
quality of an individual. Only on the basis of the comparison of these several criteria (thus
multi-objective) can a decision be made as to the superiority of one individual over another.
Then, as in single-objective problems, an order of individuals within the population can be

4/13
established from these reciprocal comparisons - multi-objective ranking. After this order has
been established the single-objective ranking methods from the subsection 3.1 can be used
to convert the order of the individuals to corresponding fitness values.

Multi-objective fitness assignment (and with it multi-objective optimization ) are such an

important aspect, that an own chapter contains the description of its different aspects. See
Chapter 7 for an extensive description of PARETO-ranking, goal attainment, sharing, and
much more.

3.3 Roulette wheel selection

The simplest selection scheme is roulette-wheel selection, also called stochastic sampling
with replacement [Bak87]. This is a stochastic algorithm and involves the following
technique:

The individuals are mapped to contiguous segments of a line, such that each individual's
segment is equal in size to its fitness. A random number is generated and the individual
whose segment spans the random number is selected. The process is repeated until the
desired number of individuals is obtained (called mating population). This technique is
analogous to a roulette wheel with each slice proportional in size to the fitness, see figure .

Table shows the selection probability for 11 individuals, linear ranking and selective
pressure of 2 together with the fitness value. Individual 1 is the most fit individual and
occupies the largest interval, whereas individual 10 as the second least fit individual has
the smallest interval on the line (see figure ). Individual 11, the least fit interval, has a fitness
value of 0 and get no chance for reproduction

Table 3-2: Selection probability and fitness value

Number of individual 1 2 3 4 5 6 7 8 9 10 11

fitness value 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0

selection probability 0.18 0.16 0.15 0.13 0.11 0.09 0.07 0.06 0.03 0.02 0.0

For selecting the mating population the appropriate number of uniformly distributed random
numbers (uniform distributed between 0.0 and 1.0) is independently generated.

sample of 6 random numbers:

0.81, 0.32, 0.96, 0.01, 0.65, 0.42.

Figure shows the selection process of the individuals for the example in table together with
the above sample trials.

Fig. 3-3: Roulette-wheel selection

5/13
After selection the mating population consists of the individuals:

1, 2, 3, 5, 6, 9.

The roulette-wheel selection algorithm provides a zero bias but does not guarantee
minimum spread.

3.4 Stochastic universal sampling

Stochastic universal sampling [Bak87] provides zero bias and minimum spread. The
individuals are mapped to contiguous segments of a line, such that each individual's
segment is equal in size to its fitness exactly as in roulette-wheel selection. Here equally
spaced pointers are placed over the line as many as there are individuals to be selected.
Consider NPointer the number of individuals to be selected, then the distance between the
pointers are 1/NPointer and the position of the first pointer is given by a randomly
generated number in the range [0, 1/NPointer].

For 6 individuals to be selected, the distance between the pointers is 1/6=0.167. Figure
shows the selection for the above example.

sample of 1 random number in the range [0, 0.167]:

0.1.

Fig. 3-4: Stochastic universal sampling

After selection the mating population consists of the individuals:

1, 2, 3, 4, 6, 8.

Stochastic universal sampling ensures a selection of offspring which is closer to what is

deserved then roulette wheel selection.

3.5 Local selection

In local selection every individual resides inside a constrained environment called the local
6/13
neighborhood. (In the other selection methods the whole population or subpopulation is the
selection pool or neighborhood.) Individuals interact only with individuals inside this region.
The neighborhood is defined by the structure in which the population is distributed. The
neighborhood can be seen as the group of potential mating partners.

Local selection is part of the local population model, see Section 8.2.

The first step is the selection of the first half of the mating population uniform at random (or
using one of the other mentioned selection algorithms, for example, stochastic universal
sampling or truncation selection). Now a local neighborhood is defined for every selected
individual. Inside this neighborhood the mating partner is selected (best, fitness
proportional, or uniform at random).

The structure of the neighborhood can be:

linear
full ring, half ring (see Figure )
two-dimensional
full cross, half cross (see Figure , left)
full star, half star (see Figure , right)
three-dimensional and more complex with any combination of the above structures.

Fig. 3-5: Linear neighborhood: full and half ring

The distance between possible neighbors together with the structure determines the size of
the neighborhood. Table gives examples for the size of the neighborhood for the given
structures and different distance values.

Fig. 3-6: Two-dimensional neighborhood; left: full and half cross, right: full and half star

7/13
Between individuals of a population an
`isolation by distance' exists. The smaller
the neighborhood, the bigger the isolation
distance. However, because of overlapping
neighborhoods, propagation of new
variants takes place. This assures the
exchange of information between all
individuals.

The size of the neighborhood determines

the speed of propagation of information
between the individuals of a population,
thus deciding between rapid propagation or
maintenance of a high diversity/variability in
the population. A higher variability is often
desired, thus preventing problems such as
premature convergence to a local minimum.
Similar results were drawn from simulations
in [VBS91]. Local selection in a small
neighborhood performed better than local
selection in a bigger neighborhood.
Nevertheless, the interconnection of the
whole population must still be provided.
Two-dimensional neighborhood with
structure half star using a distance of 1 is
recommended for local selection. However, if the population is bigger (>100 individuals) a
greater distance and/or another two-dimensional neighborhood should be used.

Tab. 3-3: Number of neighbors for local selection

distance

structure of selection 1 2

full ring 2 4

half ring 1 2

full cross 4 8 (12)

half cross 2 4 (5)

full star 8 24

half star 3 8

8/13
 3.6 Truncation selection

Compared to the previous selection methods modeling natural selection truncation

selection is an artificial selection method. It is used by breeders for large populations/mass
selection.

In truncation selection individuals are sorted according to their fitness. Only the best
individuals are selected for parents. These selected parents produce uniform at random
offspring. The parameter for truncation selection is the truncation threshold Trunc. Trunc
indicates the proportion of the population to be selected as parents and takes values
ranging from 50%-10%. Individuals below the truncation threshold do not produce offspring.
The term selection intensity is often used in truncation selection. Table shows the relation
between both.

Tab. 3-4: Relation between truncation threshold and selection intensity

truncation threshold 1% 10% 20% 40% 50% 80%

selection intensity 2.66 1.76 1.2 0.97 0.8 0.34

3.6.1 Analysis of truncation selection

In [BT95] an analysis of truncation selection can be found. The same results have been
derived in a different way in [CK70] as well.

Selection intensity

(3-7)

Loss of diversity

(3-8)

Selection variance

(3-9)

Fig. 3-7: Properties of truncation selection

9/13
 3.7 Tournament selection

In tournament selection [GD91] a number Tour of individuals is chosen randomly from the
population and the best individual from this group is selected as parent. This process is
repeated as often as individuals must be chosen. These selected parents produce uniform
at random offspring. The parameter for tournament selection is the tournament size Tour.
Tour takes values ranging from 2 to Nind (number of individuals in population). Table and
figure show the relation between tournament size and selection intensity [BT95].

Tab. 3-5: Relation between tournament size and selection intensity

tournament size 1 2 3 5 10 30

selection intensity 0 0.56 0.85 1.15 1.53 2.04

3.7.1 Analysis of tournament selection

In [BT95] an analysis of tournament selection can be found.

Selection intensity

(3-10)

Loss of diversity

(3-11)
(About 50% of the population are
lost at tournament size Tour=5).

Selection variance

10/13
 (3-12)

Fig. 3-8: Properties of tournament selection

3.8 Comparison of selection schemes

As shown in the previous sections of this chapter the selection methods behave similarly
assuming similar selection intensity.

3.8.1 Selection parameter and selection intensity

Figure shows the relation between selection intensity and the appropriate parameters of the
selection methods (selective pressure, truncation threshold and tournament size). It should
be stated that with tournament selection only discrete values can be assigned and linear
ranking selection allows only a smaller range for the selection intensity.

However, the behavior of the selection methods is different. Thus, the selection methods
will be compared on the parameters loss of diversity (figure ) and selections variance
(figure ) on the selection intensity.

Fig. 3-9: Dependence of selection parameter on selection intensity

11/13
3.8.2 Loss of diversity and selection intensity
Fig. 3-10: Dependence of loss of diversity on selection intensity

Truncation selection leads to a much higher loss of diversity for the same selection
intensity compared to ranking and tournament selection. Truncation selection is more likely
to replace less fit individuals with fitter offspring, because all individuals below a certain
fitness threshold do not have a probability to be selected. Ranking and tournament
selection seem to behave similarly. However, ranking selection works in an area where
tournament selection does not work because of the discrete character of tournament
selection.

3.8.3 Selection variance and selection intensity

Fig. 3-11: Dependence of selection variance on selection intensity

12/13
For the same selection intensity truncation selection leads to a much smaller selection
variance than ranking or tournament selection. As can be seen clearly ranking selection
behaves similar to tournament selection. However, again ranking selection works in an area
where tournament selection does not work because of the discrete character of tournament
selection. In [BT95] was proven that the fitness distribution for ranking and tournament
selection for SP=2 and Tour=2 (SelInt=1/sqrt(pi)) is identical.

This document is part of version 3.8 of the GEATbx: Genetic and Evolutionary Algorithm Toolbox for use
with Matlab - www.geatbx.com.
The Genetic and Evolutionary Algorithm Toolbox is not public domain.
© 1994-2006 Hartmut Pohlheim, All Rights Reserved, ([email protected]).

13/13
 geatbx.com/docu/algindex-03.html

Evolutionary Algorithms 4 Recombination

4 Recombination
Recombination produces new individuals in combining the information contained in two or
more parents (parents - mating population). This is done by combining the variable values
of the parents. Depending on the representation of the variables different methods must be
used.

Section 4.1 describes the discrete recombination. This method can be applied to all
variable representations. Section 4.2 explains methods for real valued variables. Methods
for binary valued variables are described in Section 4.3.

The methods for binary valued variables constitute special cases of the discrete
recombination. These methods can all be applied to integer valued and real valued
variables as well.

4.1 All representations - Discrete

recombination

Discrete recombination [MSV93a] performs an exchange of variable values between the

individuals. For each position the parent who contributes its variable to the offspring is
chosen randomly with equal probability.

(4-1)

Discrete recombination generates corners of the hypercube defined by the parents. Figure
shows the geometric effect of discrete recombination.

Fig. 4-1: Possible positions of the offspring after discrete recombination

1/9
Consider the following two individuals with 3 variables each (3 dimensions), which will also
be used to illustrate the other types of recombination for real valued variables:

individual 1 12 25 5
individual 2 123 4 34

For each variable the parent who contributes its variable to the offspring is chosen
randomly with equal probability.:

sample 1 2 2 1
sample 2 1 2 1

After recombination the new individuals are created:

offspring 1 123 4 5
offspring 2 12 4 5

Discrete recombination can be used with any kind of variables (binary, integer, real or
symbols).

4.2 Real valued recombination

The recombination methods in this section can be applied for the recombination of
individuals with real valued variables.

4.2.1 Intermediate recombination

Intermediate recombination [MSV93a] is a method only applicable to real variables (and not
binary variables). Here the variable values of the offspring are chosen somewhere around
and between the variable values of the parents.

Offspring are produced according to the rule:

(4-2)

where a is a scaling factor chosen uniformly at random over an interval [-d, 1+d] for each
variable anew.

2/9
The value of the parameter d defines the size of the area for possible offspring. A value of
d = 0 defines the area for offspring the same size as the area spanned by the parents. This
method is called (standard) intermediate recombination. Because most variables of the
offspring are not generated on the border of the possible area, the area for the variables
shrinks over the generations. This shrinkage occurs just by using (standard) intermediate
recombination. This effect can be prevented by using a larger value for d. A value of
d = 0.25 ensures (statistically), that the variable area of the offspring is the same as the
variable area spanned by the variables of the parents. See figure 4-2 for a picture of the
area of the variable range of the offspring defined by the variables of the parents.

Fig. 4-2: Area for variable value of offspring compared to parents in intermediate
recombination

Consider the following two individuals with 3 variables each:

individual 1 12 25 5
individual 2 123 4 34

The chosen a for this example are:

sample 1 0.5 1.1 -0.1

sample 2 0.1 0.8 0.5

The new individuals are calculated as:

offspring 1 67.5 1.9 2.1

offspring 2 23.1 8.2 19.5

Intermediate recombination is capable of producing any point within a hypercube slightly

larger than that defined by the parents. Figure 4-3 shows the possible area of offspring after
intermediate recombination.

Fig. 4-3: Possible area of the offspring after intermediate recombination

4.2.2 Line recombination

3/9
Line recombination [MSV93a] is similar to intermediate recombination, except that only one
value of a for all variables is used. The same a is used for all variables:

(4-3)

For the value of d the statements given for intermediate recombination are applicable.

Consider the following two individuals with 3 variables each:

individual 1 12 25 5
individual 2 123 4 34

The chosen Alpha for this example are:

sample 1 0.5
sample 2 0.1

The new individuals are calculated as:

offspring 1 67.5 14.5 19.5

offspring 2 23.1 22.9 7.9

Line recombination can generate any point on the line defined by the parents. Figure 4-4
shows the possible positions of the offspring after line recombination.

Fig. 4-4: Possible positions of the offspring after line recombination

4.2.3 Extended line recombination

Extended line recombination [Müh94] generates offspring on a line defined by the variable
values of the parents. However, extended line recombination is not restricted to the line
between the parents and a small area outside. The parents just define the line where
possible offspring may be created. The size of the area for possible offspring is defined by
the domain of the variables.

Inside this possible area the offspring are not uniform at random distributed. The probability
of creating offspring near the parents is high. Only with low probability offspring are created
far away from the parents. If the fitness of the parents is available, then offspring are more

4/9
often created in the direction from the worse to the better parent (directed extended line
recombination).

Offspring are produced according to the following rule:

(4-4)

The creation of offspring uses features similar to the mutation operator for real valued
variables (see Section 5.1). The parameter a defines the relative step-size, the parameter r
the maximum step-size and the parameter s the direction of the recombination step.

Figure 4-5 tries to visualize the effect of extended line recombination.

The parameter k determines the precision used for the creation of recombination steps. A
larger k produces more smaller steps. For all values of k the maximum value for a is a = 1
(u = 0). The minimum value of a depends on k and is a =2-k (u = 1). Typical values for the
precision parameter k are in the area from 4 to 20.

Fig. 4-5: Possible positions of the offspring after extended line recombination according to
the positions of the parents and the definition area of the variables

A robust value for the parameter r (range of recombination step) is 10% of the domain of
the variable. However, according to the defined domain of the variables or for special cases
this parameter can be adjusted. By selecting a smaller value for r the creation of offspring
may be constrained to a smaller area around the parents.

If the parameter s (search direction) is set to -1 or +1 with equal probability an undirected

recombination takes place. If the probability of s=+1 is higher than 0.5, a directed
5/9
recombination takes place (offspring are created in the direction from the worse to the
better parent - the first parent must be the better parent).

Extended line recombination is only applicable to real variables (and not binary or integer
variables).

4.3 Binary valued recombination

(crossover)

This section describes recombination methods for individuals with binary variables.
Commonly, these methods are called 'crossover'. Thus, the notion 'crossover' will be used
to name the methods.

During the recombination of binary variables only parts of the individuals are exchanged
between the individuals. Depending on the number of parts, the individuals are divided
before the exchange of variables (the number of cross points). The number of cross points
distinguish the methods.

4.3.1 Single-point / double point / multi-point crossover

In single-point crossover one crossover position k_[1,2,...,Nvar-1], Nvar: number of
variables of an individual, is selected uniformly at random and the variables exchanged
between the individuals about this point, then two new offspring are produced. Figure 4-6
illustrates this process.

Consider the following two individuals with 11 binary variables each:

individual 1 0 1 1 1 0 0 1 1 0 1 0
individual 2 1 0 1 0 1 1 0 0 1 0 1

The chosen crossover position is:

crossover position 5

After crossover the new individuals are created:

offspring 1 0 1 1 1 0| 1 0 0 1 0 1
offspring 2 1 0 1 0 1| 0 1 1 0 1 0

Fig. 4-6: Single-point crossover

6/9
In double-point crossover two crossover positions are selected uniformly at random and the
variables exchanged between the individuals between these points. Then two new offspring
are produced.

Single-point and double-point crossover are special cases of the general method multi-point
crossover.

For multi-point crossover, m crossover positions k i_[1,2,...,Nvar-1], i=1:m, Nvar: number of

variables of an individual, are chosen at random with no duplicates and sorted into
ascending order. Then, the variables between successive crossover points are exchanged
between the two parents to produce two new offspring. The section between the first
variable and the first crossover point is not exchanged between individuals. Figure 4-7
illustrates this process.

Consider the following two individuals with 11 binary variables each:

individual 1 0 1 1 1 0 0 1 1 0 1 0
individual 2 1 0 1 0 1 1 0 0 1 0 1

The chosen crossover positions are:

cross pos. (m=3) 2 6 10

After crossover the new individuals are created:

offspring 1 0 1| 1 0 1 1| 0 1 1 1| 1
offspring 2 1 0| 1 1 0 0| 0 0 1 0| 0

Fig. 4-7: Multi-point crossover

The idea behind multi-point, and indeed many of the variations on the crossover operator,
is that parts of the chromosome representation that contribute most to the performance of a
particular individual may not necessarily be contained in adjacent substrings [Boo87].
Further, the disruptive nature of multi-point crossover appears to encourage the exploration
of the search space, rather than favouring the convergence to highly fit individuals early in
the search, thus making the search more robust [SDJ91b].

4.3.2 Uniform crossover

Single and multi-point crossover define cross points as places between loci where an
individual can be split. Uniform crossover [Sys89] generalizes this scheme to make every
locus a potential crossover point. A crossover mask, the same length as the individual

7/9
structure is created at random and the parity of the bits in the mask indicate which parent
will supply the offspring with which bits. This method is identical to discrete recombination,
see Section 4.1.

Consider the following two individuals with 11 binary variables each:

individual 1 0 1 1 1 0 0 1 1 0 1 0
individual 2 1 0 1 0 1 1 0 0 1 0 1

For each variable the parent who contributes its variable to the offspring is chosen
randomly with equal probability. Here, the offspring 1 is produced by taking the bit from
parent 1 if the corresponding mask bit is 1 or the bit from parent 2 if the corresponding
mask bit is 0. Offspring 2 is created using the inverse of the mask, usually.

sample 1 0 1 1 0 0 0 1 1 0 1 0
sample 2 1 0 0 1 1 1 0 0 1 0 1

After crossover the new individuals are created:

offspring 1 1 1 1 0 1 1 1 1 1 1 1
offspring 2 0 0 1 1 0 0 0 0 0 0 0

Uniform crossover, like multi-point crossover, has been claimed to reduce the bias
associated with the length of the binary representation used and the particular coding for a
given parameter set. This helps to overcome the bias in single-point crossover towards
short substrings without requiring precise understanding of the significance of the individual
bits in the individuals representation. [SDJ91a] demonstrated how uniform crossover may
be parameterized by applying a probability to the swapping of bits. This extra parameter
can be used to control the amount of disruption during recombination without introducing a
bias towards the length of the representation used.

4.3.3 Shuffle crossover

Shuffle crossover [CES89] is related to uniform crossover. A single crossover position (as
in single-point crossover) is selected. But before the variables are exchanged, they are
randomly shuffled in both parents. After recombination, the variables in the offspring are
unshuffled in reverse. This removes positional bias as the variables are randomly
reassigned each time crossover is performed.

4.3.4 Crossover with reduced surrogate

The reduced surrogate operator [Boo87] constrains crossover to always produce new
individuals wherever possible. This is implemented by restricting the location of crossover
points such that crossover points only occur where gene values differ.

This document is part of version 3.8 of the GEATbx: Genetic and Evolutionary Algorithm Toolbox for use
with Matlab - www.geatbx.com.
The Genetic and Evolutionary Algorithm Toolbox is not public domain.
© 1994-2006 Hartmut Pohlheim, All Rights Reserved, ([email protected]).

8/9
9/9
 geatbx.com/docu/algindex-04.html

Evolutionary Algorithms 5 Mutation

5 Mutation
By mutation individuals are randomly altered. These variations (mutation steps) are mostly
small. They will be applied to the variables of the individuals with a low probability (mutation
probability or mutation rate). Normally, offspring are mutated after being created by
recombination.

For the definition of the mutation steps and the mutation rate two approaches exist:

Both parameters are constant during a whole evolutionary run. Examples are
methods for the mutation of real variables, see Section 5.1 and mutation of binary
variables, see Section 5.2.
One or both parameters are adapted according to previous mutations. Examples are
the methods for the adaptation of mutation step-sizes known from the area of
evolutionary strategies, see Section 5.3.

5.1 Real valued mutation

Mutation of real variables means, that randomly created values are added to the variables
with a low probability. Thus, the probability of mutating a variable (mutation rate) and the
size of the changes for each mutated variable (mutation step) must be defined.

The probability of mutating a variable is inversely proportional to the number of variables

(dimensions). The more dimensions one individual has, the smaller is the mutation
probability. Different papers reported results for the optimal mutation rate. [MSV93a] writes,
that a mutation rate of 1/n (n: number of variables of an individual) produced good results
for a wide variety of test functions. That means, that per mutation only one variable per
individual is changed/mutated. Thus, the mutation rate is independent of the size of the
population.

Similar results are reported in [Bäc93]and [Bäc96] for a binary valued representation. For
unimodal functions a mutation rate of 1/n was the best choice. An increase in the mutation
rate at the beginning connected with a decrease in the mutation rate to 1/n at the end gave
only an insignificant acceleration of the search.

The given recommendations for the mutation rate are only correct for separable functions.
However, most real world functions are not fully separable. For these functions no
1/5
recommendations for the mutation rate can be given. As long as nothing else is known, a
mutation rate of 1/n is suggested as well.

The size of the mutation step is usually difficult to choose. The optimal step-size depends
on the problem considered and may even vary during the optimization process. It is known,
that small steps (small mutation steps) are often successful, especially when the individual
is already well adapted. However, larger changes (large mutation steps) can, when
successful, produce good results much quicker. Thus, a good mutation operator should
often produce small step-sizes with a high probability and large step-sizes with a low
probability.

In [MSV93a] and [Müh94] such an operator is proposed (mutation operator of the Breeder
Genetic Algorithm):

(5-1)

This mutation algorithm is able to generate most points in the hyper-cube defined by the
variables of the individual and range of the mutation (the range of mutation is given by the
value of the parameter r and the domain of the variables). Most mutated individuals will be
generated near the individual before mutation. Only some mutated individuals will be far
away from the not mutated individual. That means, the probability of small step-sizes is
greater than that of bigger steps. Figure tries to give an impression of the mutation results
of this mutation operator.

Fig. 5-1: Effect of mutation of real variables in two dimensions

The parameter k (mutation precision) defines indirectly the minimal step-size possible and
the distribution of mutation steps inside the mutation range. The smallest relative mutation
step-size is 2-k , the largest 20 = 1. Thus, the mutation steps are created inside the area [r,
r·2-k ] (r: mutation range). With a mutation precision of k = 16, the smallest mutation step

-16 2/5
possible is r·2-16. Thus, when the variables of an individual are so close to the optimum, a
further improvement is not possible. This can be circumvented by decreasing the mutation
range (restart of the evolutionary run or use of multiple strategies)

Typical values for the parameters of the mutation operator from equation 5-1 are:

(5-2)

By changing these parameters very different

search strategies can be defined.

5.2 Binary mutation

For binary valued individuals mutation means the flipping of variable values, because every
variable has only two states. Thus, the size of the mutation step is always 1. For every
individual the variable value to change is chosen (mostly uniform at random). Table shows
an example of a binary mutation for an individual with 11 variables, where variable 4 is
mutated.

Tab. 5-1: Individual before and after binary mutation

Assuming that the above individual decodes a real number in the bounds [1, 10], the effect
of the mutation depends on the actual coding. Table shows the different numbers of the
individual before and after mutation for binary/gray and arithmetic/logarithmic coding.

Tab. 5-2: Result of the binary mutation

However, there is no longer a reason to decode real variables into binary variables.
Powerful mutation operators for real variables are available, see the operator in
Section 5.1. The advantages of these operators were shown in some publications (for
instance [Mic94] and [Dav91]).

3/5
 5.3 Real valued mutation with adaptation of
step-sizes

For the mutation of real variables exists the possibility to learn the direction and step-size of
successful mutations by adapting these values. These methods are a part of evolutionary
strategies ([Sch81] and [Rec94]) and evolutionary programming ([Fdb95]).

Extensions of these methods or new developments were published recently:

Adaptation of n (number of variables) step-sizes ([OGH93], [OGH94]: ES-algorithm

with derandomized mutative step-size control using accumulated information),
Adaptation of n step-sizes and one direction ([HOG95]: derandomized adaptation of n
individual step-sizes and one direction - A II),
Adaptation of n step-sizes and n directions ([HOG95]: derandomized adaptation of
the generating set - A I).

For storing the additional mutation step-sizes and directions additional variables are added
to every individual. The number of these additional variables depends on the number of
variables n and the method. Each step-size corresponds to one additional variable, one
direction to n additional variables. To store n directions n2 additional variables would be
needed.

In addition, for the adaptation of n step-sizes n generations with the calculation of multiple
individuals each are needed. With n step-sizes and one direction (A II) this adaptation takes
2n generations, for n directions (A I) n2 generations.

When looking at the additional storage space required and the time needed for adaptation it
can be derived, that only the first two methods are useful for practical application. Only
these methods achieve an adaptation with acceptable expenditure. The adaptation of n
directions (A I) is currently only applicable to small problems.

The algorithms for these mutation operators will not be described at this stage. Instead, the
interested reader will be directed towards the publications mentioned. An example
implementation is contained in [GEATbx]. Some comments important for the practical use
of these operators will be given in the following paragraphs.

The mutation operators with step-size adaptation need a different setup for the evolutionary
algorithm parameters compared to the other algorithms. The adapting operators employ a
small population. Each of these individuals produces a large number of offspring. Only the
best of the offspring are reinserted into the population. All parents will be replaced. The
selection pressure is 1, because all individuals produce the same number of offspring. No
recombination takes place.

Good values for the mentioned parameters are:

1 (1-3) individuals per population or subpopulation,

5 (3-10) offspring per individual => generation gap = 5,
4/5
 the best offspring replace parents => reinsertion rate = 1,
no selection pressure => SP = 1,
no recombination.

When these mutation operators were used one problem had to be solved: the initial size of
the individual step-sizes. The original publications just give a value of 1. This value is only
suitable for a limited number of artificial test functions and when the domain of all variables
is equal. For practical use the individual initial step-sizes must be defined depending on the
domain of each variable. Further, a problem-specific scaling of the initial step-sizes should
be possible. To achieve this the parameter mutation range r can be used, similar to the real
valued mutation operator.

Typical values for the mutation range of the adapting mutation operators are:

(5-3)

The mutation range determines the initialization of the

step-sizes at the beginning of a run only. During the following step-size adaptation the step-
sizes are not constrained.

A larger value for the mutation range produces larger initial mutation steps. The offspring
are created far away from the parents. Thus, a rough search is performed at the beginning
of a run. A small value for the mutation range determines a detailed search at the
beginning. Between both extremes the best way to solve the problem at hand must be
selected. If the search is too rough, no adaptation takes place. If the initial step sites are too
small, the search takes extraordinarily long and/or the search gets stuck in the next small
local minimum.

The adapting mutation operators should be especially powerful for the solution of problems
with correlated variables. By the adaptation of step-sizes and directions the correlations
between variables can be learned. Some problems (for instance the Rosenbrock function -
contains a small and curve shaped valley) can be solved very effectively by adapting
mutation operators.

The use of the adapting mutation operators is very difficult (or useless), when the objective
function contains many minima (extrema) or is noisy.

This document is part of version 3.8 of the GEATbx: Genetic and Evolutionary Algorithm Toolbox for use
with Matlab - www.geatbx.com.
The Genetic and Evolutionary Algorithm Toolbox is not public domain.
© 1994-2006 Hartmut Pohlheim, All Rights Reserved, ([email protected]).

5/5
 geatbx.com/docu/algindex-05.html

Evolutionary Algorithms 6 Reinsertion

6.1 Global reinsertion

6.2 Local reinsertion

6 Reinsertion
Once the offspring have been produced by selection, recombination and mutation of
individuals from the old population, the fitness of the offspring may be determined. If less
offspring are produced than the size of the original population then to maintain the size of
the original population, the offspring have to be reinserted into the old population. Similarly,
if not all offspring are to be used at each generation or if more offspring are generated than
the size of the old population then a reinsertion scheme must be used to determine which
individuals are to exist in the new population.

The used selection method determines the reinsertion scheme: local reinsertion for local
selection and global reinsertion for all other selection methods.

6.1 Global reinsertion

Different schemes of global reinsertion exist:

produce as many offspring as parents and replace all parents by the offspring (pure
reinsertion).
produce less offspring than parents and replace parents uniformly at random (uniform
reinsertion).
produce less offspring than parents and replace the worst parents (elitist reinsertion).
produce more offspring than needed for reinsertion and reinsert only the best
offspring (fitness-based reinsertion).

Pure Reinsertion is the simplest reinsertion scheme. Every individual lives one generation
only. This scheme is used in the simple genetic algorithm. However, it is very likely, that
very good individuals are replaced without producing better offspring and thus, good
information is lost.

Fig. 6-1: Scheme for elitist insertion

1/3
The elitist combined with fitness-based reinsertion prevents this losing of information and is
the recommended method. At each generation, a given number of the least fit parents is
replaced by the same number of the most fit offspring (see figure ). The fitness-based
reinsertion scheme implements a truncation selection between offspring before inserting
them into the population (i.e. before they can participate in the reproduction process). On
the other hand, the best individuals can live for many generations. However, with every
generation some new individuals are inserted. It is not checked whether the parents are
replaced by better or worse offspring.

Because parents may be replaced by offspring with a lower fitness, the average fitness of
the population can decrease. However, if the inserted offspring are extremely bad, they will
be replaced with new offspring in the next generation.

6.2 Local reinsertion

In local selection individuals are selected in a bounded neighborhood. (see Section 3.5).
The reinsertion of offspring takes place in exactly the same neighborhood. Thus, the
locality of the information is preserved.

The used neighborhood structures are the same as in local selection. The parent of an
individual is the first selected parent in this neighborhood.

For the selection of parents to be replaced and for selection of offspring to reinsert the
following schemes are possible:

insert every offspring and replace individuals in neighborhood uniformly at random,

insert every offspring and replace weakest individuals in neighborhood,
insert offspring fitter than weakest individual in neighborhood and replace weakest
individuals in neighborhood,
insert offspring fitter than weakest individual in neighborhood and replace parent,
insert offspring fitter than weakest individual in neighborhood and replace individuals
in neighborhood uniformly at random,
insert offspring fitter than parent and replace parent.

This document is part of version 3.8 of the GEATbx: Genetic and Evolutionary Algorithm Toolbox for use
with Matlab - www.geatbx.com.
The Genetic and Evolutionary Algorithm Toolbox is not public domain.

2/3
© 1994-2006 Hartmut Pohlheim, All Rights Reserved, ([email protected]).

3/3
 perlmonks.org

A Beginning Guide To Evolutionary Algorithms

by blokhead

What Are They?

In an evolutionary algorithm (EA), we search the solution space (the set of all possible inputs) of a
difficult problem for the best solutions, but not naïvely like in a random or brute-force search. We
use some biological principles to help guide the search:
Reproduction
New solutions don't just appear out of thin air (or our random number generator) -- we combine
existing good solutions to come up with new (hopefully even better) solutions.
Mutation
Just like in real life, mutations can help or hurt. But they keep the gene pool from becoming stagnant
and homogenous -- preventing EA "inbreeding".
Survival of the fittest
The "good" solutions in the population are the ones we pick to pass down their genes. We eliminate
the poor solutions from consideration.

EA is an umbrella term encompassing ideas used in the more specific fields of genetic programming,
genetic algorithms, and evolutionary computing, so the concepts covered here apply to those fields
as well.
Armed with just these principles, you could implement your own rudimentary (and working) EAs.
You may already have implemented something like this before. However, as they say, the devil's in
the details. It's important to understand how the implementation details affect your EA.

Why/When Use EAs?

(Adapted from evoweb's take on the subject)
You only need to specify the goal of the computation, and need not have a complete
mathematical understanding of the problem.
Works well with incomplete and noisy data.
Easy and cheap to implement.
Easily adaptable by simply modifying parameters.
Efficiently parallelized.

On top of all that, EAs are refreshingly fun to use compared to other forms of analysis, and they
leave plenty of room for creativity. Here are a few examples of real-world applications that I know
about:
Finding circuit layouts that minimize production costs, using graph representations.
Modelling biological principles like cooperation, speciation, specialization.
Finding effective movement strategies for cheap (and dumb) robots.
Creating classifiers and predictors for data sets, often with neural nets.
Generating music/art that satisfies certain aesthetic criteria.

The Basics:
In EAs, we work with a population of individuals, data structures that represent elements of the
problem's solution space. When we talk about representation, we mean the type of internal data
structure the EA uses to store the individuals (array, string, tree, neural net, etc). Representation
does matter, but for the scope of this document's examples, strings and arrays will suffice. As they
are easily available as native data types in any sane language, they are much easier to implement
and conceptualize. The actual encoding of a solution into a data structure is called the individual's
gene. We'll talk a little more about representation later.
Fitness:
If our goal is to find individuals which represent "good" solutions, we should probably be a little more
1/5
specific about what we mean by "good." We must have a way of scoring an individual's effectiveness
for the problem. We call this measurement the individual's fitness (as in survival of the fittest). The
fitness measure should reflect the characteristics you desire in a "good" solution, so the higher an
individual's fitness, the better it demonstrates the traits you want. The fitness measure is always
dependent on the representation you use. It sets your EA's goal.
Commonly, fitness is just a function of the individual's gene data structure. However, the fitness
measure need not be a true function in the mathematical sense. It might be probablistic, or it might
depend also on other members of the population. It also often involves a model or simulation of the
problem, executed with the individuals of the population.

The Process:
The most basic evolutionary algorithm psuedocode is rather simple:
Create an initial population (usually at random)
Until "done": (exit criteria)
Select some pairs to be parents (selection)
Combine pairs of parents to create offspring (recombination)
Perform some mutation(s) on the offspring (mutation)
Select some population members to be replaced by the new offspring (replacement)
Repeat

This is extremely general, so let's look at each step in a little more detail. As you can imagine, there
are a zillion ways to fill in the implementation details of each step, so I'll only list the most common
ones.
The exit criteria sets the target for the fitness measure, but also usually includes an upper limit on
the number of iterations, in case the evolution gets "stuck." A typical exit criteria might be: "stop
when some individual achieves a fitness of 100, or when we have iterated 10,000 times." We'll talk
more about evolution getting "stuck" later. Sticking with the biology jargon, each iteration of the loop
is called a generation.

Selection and replacement grant breeding rights and cause extinction within the population,
respectively. They are independent of the representation scheme, and should only rely on your
choice of fitness measure. Usually a small fraction of the population are chosen for breeding or
replacement each generation. For simplicity, often the same number of individuals are chosen for
breeding and replacement, although this is not required (causing the population to change in size).
Here are a few of the most common selection and replacement methods:

Random: Choose completely at random, with uniform probability given to each individual
(regardless of fitness).
Absolute: Always breed the n best-fit individuals, and replace the n least-fit individuals. (No
randomness, always a deterministic choice)
Roulette: Pick randomly, but with relative weights proportional to fitness. Higher-fit individuals
have a better chance of getting chosen for breeding, and less-fit individuals have a better
chance of getting chosen for replacement
Rank: Same as roulette, but make the relative weights proportional to an individual's rank
within the population, not fitness. The least-fit individual has rank 1, while the most-fit has
rank N (the size of the population).

Selection and replacement methods are independent of each other. You could use absolute
replacement with rank selection, for example.
Recombination (or breeding) is the process of using existing pairs of "parent" genes to produce
new "offspring" genes. The details of this operation depend on your representation scheme, but by
far the most common recombination operation is called crossover. Crossover can be used with
string and array representations. It involves making copies of the parents and then swapping a
chunk between the copies. Here's a visual example on two string genes:

parent 1: "aaaaaaaaaaaaaa" (these are only copies -- the parents a

+re
parent 2: "AAAAAAAAAAAAAA" not modified in this operation)
cut here: ^ ^ (these two points chosen at random)
and then swap sections..
result 1: "aaaaaaAAAAAAaa"
result 2: "AAAAAAaaaaaaAA"

[download]

The concept of crossover can be extended and used in other representations as well. For instance,
a crossover operation on two tree structures might involve the exchange of two subtrees. Common
variations on crossover include swapping chunks from different parts of the two genes or
exchanging more than one chunk.
2/5
Mutation is a random process which slightly modifies the gene of an individual. With string genes, a
mutation usually consists of changing a fixed number of characters, or changing each character with
a very low probability (e.g, a 5% chance of changing each character). Other interesting mutations
include lengthening, shortening, or modifying the gene, each with a respective probability.

A "Hello World" Example:

It's time for an example! The "Hello World" of EAs is called ONE-MAX. The problem is to produce a
string of all 1s, starting from a population of random binary strings. This may seem silly since we
already know the final answer, but the same could be said for "Hello World" programs in
programming languages. In this case, it's the EA process and the concepts, not the final answer,
that are most important. Using a string representation for the genes, the code is as follows:

use strict;
use List::Util qw/shuffle sum/;

my $str_length = 20;
my $pop_size = 50;

my @population = sort { fitness($a) <=> fitness($b) }

map { rand_string() } 1 .. $pop_size;

my $generations;
while ( $generations++ < 1000 and fitness($population[-1]) != $str_len
+gth ) {
my @parents = shuffle @population[-10 .. -1];

my @children;
push @children, crossover( splice(@parents, 0, 2) )
while @parents;

@population[0 .. 4] = map { mutate($_) } @children;

@population = sort { fitness($a) <=> fitness($b) } @population;

printf "Average fitness after %d generations is: %g\n",

$generations, (sum map { fitness($_) } @population)/@populatio
+n;
}

#####

sub fitness {
return $_[0] =~ tr/1/1/;
}

sub crossover {
my ($s1, $s2) = @_;
my ($start, $end) = sort {$a <=> $b} map { int rand length $s1 } 1
+ .. 2;

(substr($s1, $start, $end - $start), substr($s2, $start, $end - $s

+tart)) =
(substr($s2, $start, $end - $start), substr($s1, $start, $end -
+$start));
return ($s1, $s2);
}

sub mutate {
my $s = shift;
for (0 .. length($s) - 1) {
substr($s, $_, 1) = 1 - substr($s, $_, 1) if rand() < 0.2;
}
return $s;
}

sub rand_string {
join "" => map { rand() > 0.5 ? 0 : 1 } 1 .. $str_length;
}

[download]

Can you pick out which parts of this code correspond to the parts of the pseudocode? What type of
mutation was used (N-point or probabalistic)? What type of selection and replacement scheme were
used? What percentage of the population gets breeding rights at each generation? What is the exit
criteria? How could this code be made more efficient (there are many ways)? How could the EA
process be modularized? How much harder would this have been to write in C or Java? ;)
Now What? How Do I Choose?
3/5
You now probably have a feeling for the wide range of EA building blocks. But there are so many,
how will you choose what's best for a particular problem? What makes them different? It's time for a
little theory...
Fitness Landscapes & Diversity:
One way to think of how EAs solve problems is through hill-climbing. Think of breeding as a
process of exploring the solution space: starting with high-fitness individuals, recombination and
mutation bring new individuals into the population, whose genes are "nearby" the genes of the
parents. Selection and replacement fuel the up-hill part: the new individuals who have a higher
fitness will in turn be explored while the lower ones will eventually be discarded and so on, until you
discover individuals that have the highest fitness of all nearby individuals -- they are at the top of that
particular "hill" of fitness. Notice that "nearness" of other individuals is measured in the number of
mutations and/or recombinations needed to get from here to there. So your choice of mutation and
recombination operators determines the fitness landscape.
On one hand, hill-climbing casues EA populations is to slowly cluster near the tops of these hills as
they try to achieve maximum fitness. When most of the population's members are very close to one
another (very few mutations or crossovers apart), their genes are very similar, they have much
genetic material in common, and we say the population is not diverse. Hill-climbing is desired (we do
want to maximize fitness after all), but only in moderation. If it happens too fast, it's easy for the
whole population may become "stuck" on a small number of fitness hills that are not the highest in
the solution space. Mathematically speaking, these are local optima.

On the other hand, when the population is diverse and spread out in the landscape, you may
combine two "distant" parents to get a child somewhere in the middle, maybe on a new fitness hill.
This allows for more fitness hills to be discovered, reducing the chance of getting stuck on a local
optima.

(You may have noticed that in the ONE-MAX example, there are none of these. There's only one
fitness hill, with the string of all 1s at the top. Its fitness landscape is a 20-dimensional hypercube.
Mutation moves along one or more edges of the cube, and crossover moves to any vertex along the
subcube induced by the parents. Non-trivial problems generally have fitness landscapes that are too
complex to characterize.)

Here is how diversity is affected in general by the different operations:

The recombination-replacement process decreases diversity

When we reuse the good genes and throw out the bad genes, the population tends to approach a
stale and homogenous state. Since we only reuse genetic material already existing in the population
throw out some information, we can never create new genetic material.
Mutation is the only process that increases diversity
It's the only way new genetic material is introduced throughout the EA process. Without mutation,
there would be no way to achieve gene configurations that aren't merely rearrangements of the initial
population's genes. However, mutations that are too large or too frequent can make individuals
"jump" between fitness hills. We want mutation to move individuals a small enough distance that they
usually stay on the same fitness hill.
Stronger fitness bias in selection/replacement means faster decrease in diversity
If you select and replace individuals completely at random (no bias towards fitness), it will take a
long time for the population to become homogenous (but it will still happen eventually). On the other
hand, if you always breed the most-fit individuals and always replace the least-fit (no randomness,
100% fitness bias), then only the best-fit individuals (and their presumably high-fitness children) keep
producing children generation after generation. The high-fitness genetic material gets reused for
breeding over and over again, while the lower-fitness genetic material never has a chance to be
reused.

Population diversity is needed to ensure a many fitness hills are encountered. But eventually
diversity must be sacrificed so that good solutions can "climb the hill" to maximize fitness.
Representation Matters, Too!
Mutation and recombination (and therefore the fitness landscape) rely on your choice of
representation scheme. The representation should therefore make mutation and recombination
behave like the biological concepts they represent. For instance, a small change in an individual's
gene should make only a small to moderate change in its fitness characterstics. Likewise, combining
parts of the genes of two individuals should produce an individual that shares some of its parents'
characterstics. However, the result need not be merely an average of the parents; there may be
synergy between different parts of the genes.

4/5
In solving difficult problems with EAs, finding a good representation scheme with good
recombination and mutation operations can often be the hardest piece of the puzzle. There is no
magic advice for choosing the "right" representation, and in addition to adhering to these guidelines,
the choice must be feasible to implement.

Some Final Notes

EAs are a lot of fun, especially when you are modelling some real-world situation. Because at times
EAs seem to work like magic, it's easy to forget that they have the same limitations (in terms of
complexity) as any computer model. Do expect interesting results, but don't expect sentient digital
lifeforms to crawl out of the primordial ooze of your 1s and 0s.
I think you'll enjoy working with evolutionary algorithms, as they're a bit of a departure from classical
computation/analysis methods. Hopefully this guide will give you the background needed to have a
lot of fun tinkering with EAs. Be creative and inventive, and happy evolving!

blokhead

Update: fixed moved evoweb link (thanks atcroft)

Back to Meditations

5/5
Evolution of a salesman: A complete genetic algorithm
tutorial for Python
towardsdatascience.com/evolution-of-a-salesman-a-complete-genetic-algorithm-tutorial-for-python-6fe5d2b3ca35

Drawing inspiration from natural selection, genetic algorithms (GA) are a

fascinating approach to solving search and optimization problems. While much
has been written about GA (see: here and here), little has been done to show a
step-by-step implementation of a GA in Python for more sophisticated
problems. That’s where this tutorial comes in! Follow along and, by the end, you’ll have a
complete understanding of how to deploy a GA from scratch.

Introduction

The problem

In this tutorial, we’ll be using a GA to find a solution to the traveling salesman problem
(TSP). The TSP is described as follows:

“Given a list of cities and the distances between each pair of cities, what is the shortest
possible route that visits each city and returns to the origin city?”

1/7
Illustration of a potential solution to the TSP (By Xypron [Public domain], from Wikimedia
Commons)

Given this, there are two important rules to keep in mind:

1. Each city needs to be visited exactly one time

2. We must return to the starting city, so our total distance needs to be calculated
accordingly

The approach
Let’s start with a few definitions, rephrased in the context of the TSP:

Gene: a city (represented as (x, y) coordinates)

Individual (aka “chromosome”): a single route satisfying the conditions above
Population: a collection of possible routes (i.e., collection of individuals)
Parents: two routes that are combined to create a new route
Mating pool: a collection of parents that are used to create our next population (thus
creating the next generation of routes)
Fitness: a function that tells us how good each route is (in our case, how short the
distance is)
Mutation: a way to introduce variation in our population by randomly swapping two
cities in a route
Elitism: a way to carry the best individuals into the next generation

Our GA will proceed in the following steps:

2/7
1. Create the population

2. Determine fitness

3. Select the mating pool

4. Breed

5. Mutate

6. Repeat

Now, let’s see this in action.

Building our genetic algorithm

While each part of our GA is built from scratch, we’ll use a few standard packages to make
things easier:

import numpy as np, random, operator, pandas as pd, matplotlib.pyplot as plt

Create two classes: City and Fitness

We first create a City class that will allow us to create and handle our cities. These are
simply our (x, y) coordinates. Within the City class, we add a distance calculation
(making use of the Pythagorean theorem) in line 6 and a cleaner way to output the cities as
coordinates with __repr__ in line 12.

We’ll also create a Fitness class. In our case, we’ll treat the fitness as the inverse of the
route distance. We want to minimize route distance, so a larger fitness score is better.
Based on Rule #2, we need to start and end at the same place, so this extra calculation is
accounted for in line 13 of the distance calculation.

Create the population

We now can make our initial population (aka first generation). To do so, we need a way to
create a function that produces routes that satisfy our conditions (Note: we’ll create our list
of cities when we actually run the GA at the end of the tutorial). To create an individual, we
randomly select the order in which we visit each city:

This produces one individual, but we want a full population, so let’s do that in our next
function. This is as simple as looping through the createRoute function until we have as
many routes as we want for our population.

Note: we only have to use these functions to create the initial population. Subsequent
generations will be produced through breeding and mutation.

Determine fitness

Next, the evolutionary fun begins. To simulate our “survival of the fittest”, we can make use
of Fitness to rank each individual in the population. Our output will be an ordered list with
the route IDs and each associated fitness score.

3/7
Select the mating pool

There are a few options for how to select the parents that will be used to create the next
generation. The most common approaches are either fitness proportionate selection
(aka “roulette wheel selection”) or tournament selection:

Fitness proportionate selection (the version implemented below): The fitness of

each individual relative to the population is used to assign a probability of selection.
Think of this as the fitness-weighted probability of being selected.
Tournament selection: A set number of individuals are randomly selected from the
population and the one with the highest fitness in the group is chosen as the first
parent. This is repeated to chose the second parent.

Another design feature to consider is the use of elitism. With elitism, the best performing
individuals from the population will automatically carry over to the next generation, ensuring
that the most successful individuals persist.

For the purpose of clarity, we’ll create the mating pool in two steps. First, we’ll use the
output from rankRoutes to determine which routes to select in our selection function.
In lines 3–5, we set up the roulette wheel by calculating a relative fitness weight for each
individual. In line 9, we compare a randomly drawn number to these weights to select our
mating pool. We’ll also want to hold on to our best routes, so we introduce elitism in line 7.
Ultimately, the selection function returns a list of route IDs, which we can use to create
the mating pool in the matingPool function.

Now that we have the IDs of the routes that will make up our mating pool from the
selection function, we can create the mating pool. We’re simply extracting the selected
individuals from our population.

Breed
With our mating pool created, we can create the next generation in a process called
crossover (aka “breeding”). If our individuals were strings of 0s and 1s and our two rules
didn’t apply (e.g., imagine we were deciding whether or not to include a stock in a portfolio),
we could simply pick a crossover point and splice the two strings together to produce an
offspring.

However, the TSP is unique in that we need to include all locations exactly one time. To
abide by this rule, we can use a special breeding function called ordered crossover. In
ordered crossover, we randomly select a subset of the first parent string (see line 12 in
breed function below) and then fill the remainder of the route with the genes from the
second parent in the order in which they appear, without duplicating any genes in the
selected subset from the first parent (see line 15 in breed function below).

4/7
Illustration of ordered crossover (credit: Lee Jacobson)

Next, we’ll generalize this to create our offspring population. In line 5, we use elitism to
retain the best routes from the current population. Then, in line 8, we use the breed
function to fill out the rest of the next generation.

Mutate

Mutation serves an important function in GA, as it helps to avoid local convergence by

introducing novel routes that will allow us to explore other parts of the solution space.
Similar to crossover, the TSP has a special consideration when it comes to mutation.
Again, if we had a chromosome of 0s and 1s, mutation would simply mean assigning a low
probability of a gene changing from 0 to 1, or vice versa (to continue the example from
before, a stock that was included in the offspring portfolio is now excluded).

However, since we need to abide by our rules, we can’t drop cities. Instead, we’ll use swap
mutation. This means that, with specified low probability, two cities will swap places in our
route. We’ll do this for one individual in our mutate function:

Next, we can extend the mutate function to run through the new population.

Repeat
We’re almost there. Let’s pull these pieces together to create a function that produces a
new generation. First, we rank the routes in the current generation using rankRoutes . We
then determine our potential parents by running the selection function, which allows us
to create the mating pool using the matingPool function. Finally, we then create our new
generation using the breedPopulation function and then applying mutation using the
mutatePopulation function.

5/7
Evolution in motion

We finally have all the pieces in place to create our GA! All we need to do is create the
initial population, and then we can loop through as many generations as we desire. Of
course we also want to see the best route and how much we’ve improved, so we capture
the initial distance in line 3 (remember, distance is the inverse of the fitness), the final
distance in line 8, and the best route in line 9.

Running the genetic algorithm

With everything in place, solving the TSP is as easy as two steps:

First, we need a list of cities to travel between. For this demonstration, we’ll create a list of
25 random cities (a seemingly small number of cities, but brute force would have to test
over 300 sextillion routes!):

Then, running the genetic algorithm is one simple line of code. This is where art meets
science; you should see which assumptions work best for you. In this example, we have
100 individuals in each generation, keep 20 elite individuals, use a 1% mutation rate for a
given gene, and run through 500 generations:

Bonus feature: Plot the improvement

It’s great to know our starting and ending distance and the proposed route, but we would be
remiss not to see how our distance improved over time. With a simple tweak to our
geneticAlgorithm function, we can store the shortest distance from each generation in a
progress list and then plot the results.

Run the GA in the same way as before, but now using the newly created
geneticAlgorithmPlot function:

6/7
Sample output from the geneticAlgorithmPlot function

Conclusion
I hope this was a fun, hands-on way to learn how to build your own GA. Try it for yourself
and see how short of a route you can get. Or go further and try to implement a GA on
another problem set; see how you would change the breed and mutate functions to
handle other types of chromosomes. We’re just scratching the surface here!

7/7