Procedia Computer Science

Volume 51, 2015, Pages 316–325

ICCS 2015 International Conference On Computational Science

Genetic Algorithm using Theory of Chaos

Petra Snaselova and Frantisek Zboril
Faculty of Information Technology
Brno University of Technology
Brno, Czech republic
isnaselova@fit.vutbr.cz, zboril@fit.vutbr.cz

Abstract
This paper is focused on genetic algorithm with chaotic crossover operator. We have performed
some experiments to study possible use of chaos in simulated evolution. A novel genetic algo-
rithm with chaotic optimization operation is proposed to optimization of multimodal functions.
As the basis of a new crossing operator a simple equation involving chaos is used, concrete the
logistic function. The logistic function is a simple one-parameter function of the second order
that shows a chaotic behavior for some values of the parameter. Generally, solution of the logis-
tic function has three areas of its behavior: convergent, periodic and chaotic. We have supposed
that the convergent behavior leads to exploitation and the chaotic behavior aids to exploration.
The periodic behavior is probably neutral and thus it is a negligible one. Results of our experi-
ments confirm these expectations. A proposed genetic algorithm with chaotic crossover operator
leads to more efficient computation in comparison with the traditional genetic algorithm.
Keywords: optimization, genetic algorithm, chaos

1 Introduction
Chaos is the science of surprises and unpredictability. It teaches us to expect the unexpected.
While the most traditional sciences deal with supposedly predictable phenomena like are gravity,
electricity, or chemical reactions, chaos theory deals with nonlinear problems whose behaviors
are impossible to predict or control, like turbulent fluid flow, global weather patterns, healthy
heart rhythms, DNA coding sequences, and so on. However, complex dynamic systems with
complicated feedback paths may often be understood in a sense of the chaos theory. Therefore,
it is natural to consider that evolutionary processes can also be understood in a sense of the
chaos. To explore this idea we incorporated a simple logistic function into standard genetic
algorithm.

316 Selection and peer-review under responsibility of the Scientific Programme Committee of ICCS 2015

c The Authors. Published by Elsevier B.V.
doi:10.1016/j.procs.2015.05.248
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

2 Multimodal Function
A large number of real-world problems can be considered as multimodal function optimization
problems. Objective functions may have several global optima, that is, each objective function
may have the same values in several different points in the explored state space. Moreover,
it may have some local optima points in which objective function values lay nearby a global
optimum. This paper is focused on multimodal problems with one global extreme only and more
local extremes, because for solving problems with multiple global extremes further modification
of proposed algorithm is needed.

3 Genetic Algorithm
Several methods have recently been proposed for solving the multimodal optimization problems.
They can be divided into two main categories: deterministic and stochastic (metaheuristic)
methods. Deterministic methods, e.g. gradient descent method or quasi Newton method, when
they solve complex multimodal optimization problems, may easily trapped into some local
optimum as a result of deficiency in exploiting local information. They strongly depend on a
priori information about the objective function that can yield to fewer reliable results.
Stochastic algorithms combine randomness and rules mimicking several phenomena. These
phenomena include physical processes (e.g. simulated annealing proposed by Kirkpatrick [11]),
evolutionary processes (e.g. evolutionary algorithm proposed by Koza [8], de Jong [7], and Fogel
[3], and genetic algorithms (GAs) proposed by Goldberg [4] and Holland [5]), immunological
systems (e.g. artificial immune systems proposed by de Castro [2]), electromagnetism-like
(proposed by Birbil [1]) and gravitational search algorithm (proposed by Rashedi [10]).
Genetic algorithms are inspired by Darwin’s theory of biological evaluation as an optimiza-
tion algorithm. The main idea of this technique is derived from natural evolution, so there
are biological operators such as crossover, mutation and selection. Genetic algorithm has three
randomly generated phases: Initial population of chromosomes, crossover operator and muta-
tion operator [6]. Each chromosome represents some “solution” of the problem and its quality is
done by the value of so called object/fitness function. Genetic algorithm starts by generating of
some random solutions which are marked as an initial population. In the second step, random
crossovers lead to produce new offspring and in the third step, with random value of mutation a
few of genes in chromosome are changed or replaced. The new generation of candidate solutions
is then used in the next iteration of the algorithm.
Traditional GAs usually well work for single optimum problems but they fail if they have
to find multiple solutions.

4 Chaos
Very roughly, chaos is defined as a chaotic behavior of a non-linear dynamic system that is
very sensitive on initial conditions and that is described by deterministic algorithm. Nowadays,
practical use of the chaos significantly increased. Many applications in different fields have
employed chaos theory and exploited its benefits.
Implementation, experiments and analysis of optimization processes based on chaos pre-
sented in this paper show that the chaos features are important and they can improve the
efficiency of the genetic algorithm.

317
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

Chaotic function which we use in modified genetic algorithm is very well known logistic
function (1).

zn+1 = λzn (1 − zn ) (1)

This function takes a value zn , which can be any number between 0 and 1, and changes
it into a new value zn+1 . The new value of z is then back into the formula, and so on. Such
process is called an iterative process. λ parameter in equation (1) is a number between 0 and
4. It is kept constant while the iteration process is run. The value of this parameter determines
behavior of the variable z: it can be convergent, periodic or chaotic. Values of λ, that are
smaller than 3, lead to convergent solution of equation (1), i.e. behavior of z variable converges
to some constant value. When λ is between 3 and 3.56, then periodic behavior occurs. Firstly
two different values of variable z alternate (for value of λ near to 3), when value of the λ
increases, then four, eight, sixteen etc. values of z variable alternate. The solutions of equation
(1) for λ between 3.56 and 4 become fully chaotic: neither convergent nor periodic, but variable
with no discernible pattern.

4.1 The Logistic Function as a Genetic Operator

The chromosome is slightly different from traditional GA chromosome. One chromosome of
the proposed algorithm is composed of three parts solution, value of λ and mask of uniform
crossover. First, we will explain how the modified GA works, and then we will show it on a
practical example.
Part of the chromosome containing the solution encodes the solution of the optimization
function which is solved. The number of bits of solution is set by the required accuracy of
the solution (the number of decimal ciphers). The solution encodes k values representing
the coordinates x1 , x2 , . . . xk , therefore requires k -times the number of bits. The λ value in
chromosome represents the λ from equation (1) and the number of bits required for this value
is calculated similarly as numbers of bits of solution, but the number of bits is needed only
once. Uniform mask crossover contains mask for crossing and its number of bits is set to the
same number of bits as the solution. It is necessary to encode the mask into the Gray code.
The Gray code, also known as reflected binary code, is a binary numeral system where two
successive values differ in only one bit (binary digit). The problem with natural binary codes
is that together with change of one bit the number is not change only about number one, but
about more count of number. Relation between Gray code and natural binary code is shown
in figure 1.

 


 










Figure 1: Gray Coding: (a) Converting from Natural Binary to Gray, (b) Converting from Gray
to Natural Binary.

For example 3D function can be solved by chromosome, where the solution is encoded at 22
bits (value of x1 at 11 bits and value of x2 also at 11 bits), λ at 6 bits, and mask is encoded at
22 bits.

318
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

In the proposed algorithm uniform crossover is used. Now we will describe the emergence
of the first offspring, the second offspring is created in the same way, but by swapping the first
and the second parents. Each gene in the solution of the first offspring is created by copying the
corresponding gene from the solution part of first or second parent chosen according to binary
crossover mask. Where there is 1 in the crossover mask, the gene is copied from the first parent,
and where there is 0 in the mask the gene is copied from the second parent.
Bits on position of λ value are inherited from the first parent, bits of masks are created
using the modified mask of the first parent.
Mask is modified by the logistic function (1), using the λ value stored in the chromosome
and mask itself as the input. The value of xz is obtained by interpretation of the mask bits
as real value, scaled into the range from 0 to 1. New mask is set as a xz+1 to the same bit
interpretation of equation (1).
This modified crossover operation leads to self-directed evolutionary search the space for
solutions via theory of chaos.
Both λ and the solution bits are also subjected to ordinary bit-flipping mutation. Figure 2
shows a flowchart of the modified algorithm, including the new crossover operator. In table 1
is a summary of used representation operators.

     
 
  

 
      



 

  

  

Figure 2: Comparison of basic genetic algorithm flowchart with chaotic genetic algorithm
flowchart. Operation, in which the modified algorithm differs from the basic algorithm, is
highlighted.

4.2 The Use of New Crossover Operator in a Practical Example

Now, we will demonstrate the use of the new crossover operator in a practical example, the
summary of calculations and transformations from this example are shown in table 2. An
accuracy of one decimal will be sufficient for this illustrative example. λ value lies in the interval
from 0 to 4. Suppose, the number of required bits for λ value is 6 (the distance between two
neighboring points is 0,0625 in this presentation). Interval of the searched solution is given
together with an definition of optimization task [9]. Now is supposed a test area restricted to
square x1 ∈ −10, 10 and x2 ∈ −10, 10. The number of bits needed to encode the mask is
identical as the number of bits needed for the solution, both are encoded in 16 bits. The mask
is decoded into interval (0, 1).

319
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

    
      





          


      !    "!  

  # "# "$

Table 1: Table describes the contents and sizes of the genes in a chromosome, the range of
alleles and the genetic operations used on that gene. The solution represents the value of x1 ,
x2 , x3 ..., xk defined in the domain.

Suppose a randomly generated chromosome 0101010101101010 | 111001 | 1011001110010110.

The first part of the chromosome is a solution, the values of x1 and x2 . The second part of
the chromosome is the λ value and the last part of the chromosome represents the mask of
uniform crossover. It is important that the mask is encoded in Gray code. This chromosome
will be marked as the first parent, randomly generated chromosome 1110000111001101 | 101101
| 1011001001100001 as the second parent.
The first step in the proposed crossover is the crossing with mask of a solution chromosome.
Thus, after this step, we have a solution part of offspring chromosome 0101000101001011. This
step is shown in table 2a. For the second step it is necessary to decode the λ value and mask.
Decoding the mask from the binary vector v from Gray code into the interval (0, 1) in system
of real numbers is done as follows (summary in table 2b).

1. The binary vector v in Gray code is converted to binary vector v  in binary code.
2. The binary vector v  is converted to the decimal system. (k represents the number of bits
of the converted value in binary code, it means the number 8. The mask is encoded in 8
bits.)1

k
 i
m = v  · 2k−i (2)
i=1

3. Then the corresponding real number is found by the application of a linear transformation
as follows. Here, a symbol m is the calculated mask value, a is the lower limit of the
interval of m, b is the upper limit of the interval.

b−a
m=a+ · m (3)
2k − 1
New mask value is obtained by substituting the calculated mask value m for the unknown
variable xn in the Eq. (1). The newly acquired mask value is then converted back to binary
vector stored in the Gray code and together with the λ value is copied into the offspring. λ
value is decoded in a similar manner and it is substituting for the parameter λ in the Eq. (1)
during the calculation of the new mask.
1 The mask consists of a mask for x and a mask for x . The vector v thus contains only half of the all mask
1 2
binary string, it means 8 bits.

320
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

The mask in Gray code for x1 10110011 and for x2 10010110 is written as binary numbers
11011101 and 11100100 and as 221 and 228 in decimal numbers. After the transformation into
the interval (0, 1) the mask is equal to the value 0.8667 and 0.8941. This mask value after the
modifying by the Eq. (1) is for x1 and x2 0.3374 and 0.2765, after the reverse transformation
from real number to decimal system is 86 and 70, in binary code 1010110 and 1000110 and
in Gray code 1111101 and 1100101. The summary of mask crossing is shown in table 2b.
By crossing the first and second parent we will get offspring: 0101000101001010 | 111001 |
11111011100101.

Solution x1 x2 Mask x1 x2
1. parent solution 01010101 01101010 Mask in Gray code (G. c.) 10110011 10010110
2. parent solution 11100001 11001100 Mask in binary code (b. c.) 11011101 11100100
Mask of 1. parent 10110011 10010110 Mask in decimal 221 228
New offspring 01010001 01001010 Mask in interval (0.1) 0.8667 0.8941
Mask modified by (1) 0.3374 0.2765
Modified mask in decimal 86 70
Modified mask in b. c. 1010110 1000110
Modified mask in G. c. 1111101 1100101
(a) New offspring (b) Mask crossing

Table 2: Summary of the example. In (a) there is a example of crossing with mask of a solution
chromosome. From solutions of 1. and 2. parent is created new offspring. In (b) is summary
of the outputs of the proposed crossing method of a mask.

4.3 Expected Effect of New Genetic Operator

Consider the behavior of the logistic function in dependence of λ value. Suppose the λ value
is below 3. After iteratively applying the logistic function from some initial value x0 results
converge to some value xm . Speed of convergence decreases and the value of xm increases with
λ value approaching value 3. For the λ value which is larger than 3, the course of the function
is an oscillating. This will tend to produce crossover masks that preserve the higher order
bits of the mask, but vary the lower order bits. For this reason, it is desirable to encode the
chromosomes using a Gray code, because here two successive values differ in one bit (binary
digit) only. Near convergence, the mask will become fixed. Thus an individual with convergent
λ will tend to produce offspring with progressively more rigid crossover masks. Individuals with
non convergent λ will tend to have a high degree of variability in the crossover masks of their
descendants, with the variability increasing as λ approaches the maximum value of 4.
The proposed algorithm has a large influence on the quality of behavior. Low values of λ
should favor exploitation of good masks and large values of λ should favor exploration of the
space of possible masks and consequently also solutions. Therefore one should expect greater
exploration with new operator than without. Similarly, one should observe increased exploration
when λ is in the chaotic domain rather than it is in the convergent one.

5 Genetic Algorithm Using Chaos and Benchmarks

We implemented the genetic algorithm using chaos based on simple logistic function Eq. (1)
in Java and tested it using some famous benchmarks such as Easom function (this function

321
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

has one global significant extreme), Goldstein-Price function, Rotated hyper-ellipsoid (these two
function are tasks with one slight global extreme), Rastrigin function, Michalewicz function,
Schwefel function and Shubert function (functions with many local and one global extreme)
[9]. Based on the results of tests was found that the novel algorithm takes effect only when
we solve problems with multiple local extrema or problems with one global significant extreme.
The results in the remaining groups of tasks were almost the same as the results of the classic
genetic algorithm. Two functions that were chosen for representation in this publication are
defined in [9] (Easom function and Michalewicz function) and have many good features for
optimization problems such as kind of acceptable optimums.

6 Results of Solving Multimodal Problems

Now we show the results of some test that we have done with the proposed and the classical
algorithm. Parameters of GA (such as probability of crossover, etc.) that were used during
experimentation are as follows.

• Probability of crossover was 0.85,

• probability of mutation was 0.01,
• elitism was used in both algorithm,
• basic GA uses two pointed crossover and flip bit mutation,
• GA with logistic function uses new genetic crossover operator and flip bit mutation,
• calculation was performed for many generations and was ended by generation convergence,
• each run was repeated ten times and the results were averaged,
• 4 decimal places of accuracy were used.

To determine which scaling and selection algorithms give the best performance, we compared
several types of selection algorithms, namely linear and exponential arrangement, probability
roulette, tournament and probability tournament. Tournament selection became the most
successful type of selection. It is versatile and can be used for all tasks. So we used tournament
selection during further testing.
Figure 3 in 1. column shows the average of results of basic GA and proposed GA using chaos
in a) functions with one significant extreme b) functions with one slight extreme c) functions
with many local extremes. The results of experiments are the average of all task results belonging
to the three individual type of functions mentioned above. We also compare the efficiency of the
algorithms from the perspective of convergence rate depending on the size of the population.
As it is shown in the figure 3 in 1. column, the results of GA using chaos are much more efficient
than basic GA in cases c) functions with many local extremes and a little in a) functions with
one significant extreme. The differences of results of b) functions with one slight extreme are
small.
During further testing, we focused on the task with many local extremes, because they were
significantly improved. Especially we perform experiments on Michalewicz function.
The figure 3 in 2. column shows the evolution of populations in dependence on the random
numbers distribution during the initial generation of individuals. In the figure 3 a) individuals
are generated according to the uniform distribution, the b) according to the normal (Gaussian)

322
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

distribution, and the c) according to λ parameter logistic equation Eq. (1), where the same
number of individuals is generated in each interval (0, 3), (3, 3.56) and (3.56, 4). As can be seen
from the figure 3, the different distributions have no effect on the result.
(1) (2)
a) 80 100
GA chaos
Mean of iteration

GA basic 80

Population size
70
60 convergence
60 bifurcation
40 chaos
50 20
40 0
100 150 200 250 300 350 400 0 25 50 75 100
Population size Mean of iterations

b)100 100
GA chaos
Mean of iteration

90 GA basic Population size 80

80 60 convergence
bifurcation
70 40 chaos
60 20
50 0
200 250 300 350 400 0 25 50 75 100
Population size Mean of iterations

c) 90 80
GA chaos convergence
Mean of iteration

80 GA basic bifurcation
Population size

60 chaos
70
60 40
50
20
40
30 0
50 100 150 200 250 300 350 400 0 25 50 75 100
Population size Mean of iterations

Figure 3: In (1) the figures show the decreasing of time complexity when population size is
increased on: a) functions with one slight extreme, b) functions with one significant extreme,
c) funtion with many local extremes. Only solutions that converges to the correct results are
drawn. In (2) figures shows the number of specific λ values in each category (λ converging,
bifurcations and chaos) in dependence on the number of iterations of the run. The initial pop-
ulation in the picture a) was generated by a uniform distribution, in b) by a normal (Gaussian)
distribution and in c) in dependence on the boundary interval values of the parameter λ of Eq.
(1), so that to each category of parameter λ belonged the same number individuals.

To get a better handle on the effect of new crossing operator, we forced several distributions
of λ value on the initial population: including convergent values, bifurcation λ values, and

323
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

chaotic values. Different distributions are shown in the table 3. All three series of tests were
performed with the population of 200 individuals and 100 of them survived to next generation.
All three series of tests were performed with two mutation probabilities 0.01 and 0.001. Note
that the lowest number of iterations and the lowest deviation for these tests occurred for one of
the mix distribution cases. Apparently, a mixture of initial λ values in the population improves
the effectiveness of proposal algorithm. There is a weak suggestion that the proposal algorithm
with a mixed initial population is beneficial for obtaining low deviation per low number of
iterations.
Even if a mixture of λ values in the initial population is beneficial, one category of λ values
might tend to produce the highest scoring individual more frequently than the others do. As
figure 3 in 2. column presents, it might have been expected that whatever value of chaotic λ
have during a calculation tend to die off by the final generation. Individuals which possess a
chaotic λ value have a positive effect for the beginning of searching space, individuals with a
convergent λ value are beneficial at the end of calculation. Benefit of bifurcation λ values is
neutral.

Type Initial Crossing Probability Deviation Iterations

of GA population operator of mutation
Classic GA classic 2 point crossing 0.001 6.0492E-06 62
Classic GA classic 2 point crossing 0.01 6.0492E-06 51
Proposal GA mix λ crossing 0.001 6.0492E-06 34
Proposal GA mix λ crossing 0.01 6.0492E-06 36
Proposal GA convergent λ crossing 0.001 1.2166E-03 100
Proposal GA convergent λ crossing 0.01 1.9918E-04 99
Proposal GA bifurcation λ crossing 0.001 1.9918E-04 94
Proposal GA bifurcation λ crossing 0.01 2.2777E-04 66
Proposal GA chaotic λ crossing 0.001 5.5243E-05 72
Proposal GA chaotic λ crossing 0.01 5.5243E-05 53

Table 3: Comparison of different GA. Mix initial population randomly select λ from all domains.
Convergent, bifurcation and chaotic initial population selects from convergent, bifurcation and
chaotic domains. Classic initial population randomly generates population without using chaos.

7 Conclusion
In this paper, we studied using of chaos in the optimization problems. A number of interesting
trends were detected by a data.
In summary, the results show that the introduction of chaos into genetic algorithms can be
useful.
New genetic operator with a mixture of λ categories improved the rate of calculation, but
on deviation in solution it had no effect. It is only important that the initial population is
represented by the chaotic and by the convergent values. The chaotic values presumably aid
exploration, and the convergent values aid exploitation. There must be benefit in mixing of
chaotic and non-chaotic individuals. The mixing of genetic material from exploring (chaotic)
individuals and exploitative (convergent) individuals is the greatest advantage. Bifurcation of
individuals are not useful neither for exploration, nor exploitation, but merely cyclic repetition
of previously used genetic material.

324
Genetic Algorithm using Theory of Chaos Snaselova, Zboril

We also found that the different methods of distribution of initializing an initial population
does not affect the result.

Acknowledgment
This work was supported by the European Regional Development Fund in the IT4Innovations Centre
of Excellence project (CZ.1.05/1.1.00/02.0070) and by the Reliability and Security in IT project (FIT-
S-14-2486).

References
[1] Ş İlker Birbil and Shu-Chering Fang. An electromagnetism-like mechanism for global optimization.
Journal of global optimization, 25(3):263–282, 2003.
[2] Leandro Nunes De Castro and Fernando José Von Zuben. Artificial immune systems: Part i–basic
theory and applications. Universidade Estadual de Campinas, Dezembro de, Tech. Rep, 210, 1999.
[3] L.J. Fogel, A.J. Owens, and M.J. Walsh. Artificial intelligence through simulated evolution. Wiley,
Chichester, WS, UK, 1966.
[4] David E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-
Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1st edition, 1989.
[5] John H. Holland. Adaptation in Natural and Artificial Systems. MIT Press, Cambridge, MA,
USA, 1992.
[6] J. Hynek. Genetické algoritmy a genetické programování. Pruvodce (Grada). Grada, 2008.
[7] Rafal Kicinger, Tomasz Arciszewski, and Kenneth De Jong. Evolutionary computation and struc-
tural design: A survey of the state-of-the-art. Computers & Structures, 83(23):1943–1978, 2005.
[8] John R Koza. Genetic programming: A paradigm for genetically breeding populations of computer
programs to solve problems. Stanford University, Department of Computer Science, 1990.
[9] Marcin Molga and Czesław Smutnicki. Test functions for optimization needs. Test functions for
optimization needs, 2005.
[10] Esmat Rashedi, Hossein Nezamabadi-Pour, and Saeid Saryazdi. Gsa: a gravitational search algo-
rithm. Information sciences, 179(13):2232–2248, 2009.
[11] C. D. Gelatt S. Kirkpatrick and M. P. Vecchi. Optimization by simulated annealing. American
Association for the Advancement of Science, 220(4598):671–680, May 1983.

325