A Study on Evolving Self-organizing Cellular Automata based on Neural Network Genotypes

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 495
A Study on Evolving Self-organizing Cellular Automata based on Neural
Network Genotypes
Eric Mervin Anandraj1, Ahnaf Rehan Shah2
1,2U.G. Student, Dept. of Computer Science Engineering and Technology, Bennett University, Greater Noida, Uttar
Pradesh, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - This review discusses Wilfried and Istaváns’ideaof
using Artificial Neural Networks for designing self-organizing
systems. Although a lot of researchers have been able to write
functions for regenerating simple shapes such as flags and
squares, it has been difficult to generate a function torecreate
complex images and paintings. We agree with the researchers
that using a genotypic template for the cellsintheautomation
is something that should be further researched for generating
self-organizing multicellular systems having the capability to
recreate complex images. This researchanditsexperimentsdo
not provide concrete results but provide us with sufficient
evidence that the evolutionary approach especially using
Artificial Neural Network into Cellular Automata has a lot of
potential uses and should be further researched.
Key Words: cellular automata, artificial neural network,
self-organizing systems, evolutionary algorithm
1. INTRODUCTION
For the last two to three decades a lot of systems have been
identified as having self-organizing properties. Self-
organizability has been identified not only in the domain of
computer science but they were first observed in fields of
natural and social science such as physics, chemistry and
most extensively in the domain of biology. One naturally
occurring example of self-organizing can be the
crystallization of chemicals/liquids when frozen at very low
temperatures or exposed directly to a gas (external factors).
Self-organizability can be explainedastheabilityofa system,
either natural or artificial, to adapt and modify its internal
structure in response to the system’s external factors. In the
technical field, it is one of the fundamental concepts of
Systems Science and such systems are called self-organizing
systems (SOS). Elements in such systems not only change
their behaviour but also have the power to manipulateother
elements in the system for the sake of maintaining the
stability of either the system’s structure or the function of
the whole against external fluctuations.[2] These systems
have self-repairing capabilities which for a long time has
been only seen in the field of natural science.
W. Ross Ashby’s experiments which led to his book on the
foundations of cybernetics[1] are the main reason for the
modern idea of self-organizability and the recent studies on
self-organizing systems. Although this concept has been
identified in various systems, both natural and artificial, the
fundamental questions ofhowandwhyarestill unanswered.
Due to this uncertainty still being prevalent, there are still a
lot of active researches happening in this field.
Scientists up until now, have been able to identify many
examples of self-organizing systems in various fields
exhibiting the aforementioned characteristics and have
studied their mechanism in detail but understanding the
design of such systems remains a critical challenge. Recent
results however indicate that the evolutionary approach to
designing self-organizing systems to be very promising.
2. History of Cellular Automata and Self-Organizing
System
Cellular automata (sing. Cellular Automaton) also known as
cellular spaces is a system of coloured cells on a grid of
usually a specified shape that behaves according to a
particular set of rules based on the state of neighbouring
cells. The system evolves through several discretetimesteps
while the rules can be applied iteratively for as many time
steps as desired. This concept was firstdiscoveredandstudy
on by two scientist-friends Stanislaw Ulam and John von
Neumann in the 1940s.
2.1 Game of Life
One of the first examples of cellular automata and the one
that presumably popularized this concept is “Game of Life”
[8] (or simply “Life”) developed by John Horton Conway in
the year 1970. Similar to every other cellular automaton, the
cells of this system also have only two possible states, either
dead or alive. Through each generation of the system, the
cells are objected to the following 3 simple rules:
1. Survival: a particular cell can survive for the next
generation/round only if it is surrounded by 2 or 3 alive
neighbours.
2. Birth: a new cell is born or in other words, a dead cell
becomes alive in the next generation if it is surrounded by
exactly 3 alive cells.
3. Death: an alive cell can die because of one of two reasons :
(a) Overpopulation: an alive cell dies in the following
generation if it is neighboured by 4 or more alive cells.

(b) Loneliness: an alive cell can also die in the next
generation if it is not at all surrounded or surrounded by
only one alive cell.
2.2 Discovery of Self-Organization
Although the idea of self-organization can be dated back to
as early as Greek philosophy and Buddhism, the term was
first introduced in the year 1947 by Ashby[11]. Haken
further researched in this field and concluded that for any
system there is a deep associationbetweenself-organization
and selection. He also noticed that the different forms of
collective behaviour are competing against oneanother,and
compared it with Darwin’s Survival of the fittest, calling it
the Darwinian selection in the artificial world.
Recently, self-organizing systems are the main points of
discussion in all almost all the branches of science including
engineering where engineers are starting to see its
applicability[10] in relation to the emergence of nano-scale
applications and the increasing complexity of human
artefacts.
2.3 Self-organization in Game of Life
Conway’s system allows us to understand how a cellular
automaton behaves under a certain set of rules and also how
they exhibit self-organizability in the system. For example,
when considering the R-pentomino pattern in the Game of
Life system it can be noticed that structures like boats, ships,
loaf etc are generated solely through the self-organization of
the system. The R-pentomino pattern continues through
many generations generatingandconsumingsuchstructures
along the way to finally stabilize at generation 1103.
Fig. 1. R-pentomino stabilized after 1103 generations
3. Evolutionary Programming for designing Self-
Organizing Systems
Designing a self-organizing systemhasalwaysprovedto bea
challenging task for scientists and researchers even after
extensive study on the structure and functioning of these
systems. Since there is no straightforward wayfordesigning
a system with local rules where the system as a whole
exhibits self-organizing behaviours.[9]
Researchers resorted to adapting methods inspired by the
naturally occurring self-organizing systems[5]. However,
self-organizing systems designed based on such inspiration
deliver promising results, it requires the existence and
discovery of a naturally occurring system that answers the
exact same problem at hand.
Table 1. Evolutionary vs Traditonal Approach for designing a Cellular Automata
Another approach, the one generally used,wastosubject the
systems to trial and error. But these systems often exhibited
contradictoryrelationships betweenthe emergentbehaviour
and the local rules[5]. Thus there was a need for scientiststo
look into any other possible approach, where there was no
need for any knowledge of any existing self-organizing
system.
One such promising approach was the use of neural
networks to evolve the cells and their local interactions, a
method first proposed by Wilfried Elmenreich and István
Fehérvári. Their approach was more on the side of
incorporating genetic algorithms in the system.[6]
Evolutionary Approach Traditional Approach
Supports discontinuous objective function and evolutionary
multimodal optimization
Supports linear programming problems and evolutionary
singular strategies and optimization
Based on the concept of transitional probabilities Based on deterministic system
Moves from one point to next point randomly without the
current state being dependent on the past state
Moves from one point to next point in search space linearly
and no randomness is involved
Fitness function is usedtoguidesimulationstowardsoptimal
design solutions
Information derived frompreviousisusedtofindtheoptimal
design solution
Parallel search algorithms such as Breadth-first searches is
used in the particular search space
Sequential search algorithms such as LinearSearchisused in
the particular search space

Evolutionary Algorithms are searchalgorithmsthatfunction
based on the population. and each algorithm has its own
number of approaches. The initial population of the
arrangement must have a portrayal in finding the solution.
The computational complexity of the issue relies upon the
size of the underlying populace.
Fig. 2. Evolutionary Computation Cycle
The evolutionary activities, for example, hybrid and
transformation are applied iteratively in the evolutionary
algorithms until a halting condition is fulfilled. The halting
models might be utilized to end when an adequate
arrangement has been found or when there is no
improvement over various sequential ages. Evolutionary
Computation Algorithms varies from the typical Classical
techniques in many ways[7] (see Table 1).
Several potential conventional toolshavesincebeenfounded
based on the evolutionary approach which provides
increased adaptability, optimization, and scope of solutions.
Traditional techniques try to provide one solution which is
the most optimal, whereas evolutionary algorithms include
and can provide various potential solutions.
4. Model proposed by Wilfried and Istvan
P. Bentleyand S. Kumar during their research[3]noticedthat
when a cellular automaton that is to be evolved becomes
more complicated, the evolutionary approach results in
various problems such as disruption of inheritance, the
method not being able to find the solution to the problem,
among others. In technical evolution, there is 1-1 mapping
from genotype to phenotype, which is not the case in natural
evolution where a single cell can evolve into complex
systems. Thus there was a need in cellular automata to
introduce genotype descriptions from which complex self-
organizing cellular automata could emerge.
Keeping all of this in mind, Wilfried and István developed a
model consisting of one genotypical controller that is
responsible for the state and transition of each cell in the
system[4]. The state of each cell is an instance of this
controller and the algorithm responsible is realised with the
help of a simple artificial neural network (ANN) which is
modelled as a time-discrete and recurrent ANN.
4.1 Structure of the Cellular Automaton Model
Similar to other cellular automata, this model also consistsof
a grid containing cells, but the dimensions of the model are
exactly the same as the reference image. The reference
images were often small and ranged from 50-500 pixels in
size along one edge.
But unlike Conway’s Game of Life, this model has the ability
to display colour. A scale is developed from the colours ofthe
reference images by assembling a 24-bit colourcodeforeach
colour in where the bits of the 3 channels, Red, Green and
Yellow are arranged from most to least significant for their
respective channel.
In this model, each cell of the system is under the control of
the artificial neural network where all the neurons are
connected to each other and also to itself. All theconnections
between neurons has an allocated weight (w ∈ R) whereas
every neuron has a particular bias.
For every generation, an ith neuron builds over the sum of its
bias bi and weights wji of the input connections multipliedby
current outputs of the neurons j = 1,2,...n feeding the
connections. A neuron’s output in the next generation can be
determined by using the activation function F:
The artificial neural network of each cell is madeup of a total
of 20 neurons. They are divided into 5 output, 6 hidden and9
input neurons. The reason for selecting hidden neurons
according to Wilfried and István is to find a balance between
the neural network’s capability and reducing the search
space. Out of the 5 output neurons,onedeterminesthecolour
of the cell and the other output neurons and 4 pairs of the
input neurons connect with the neighbouring cells artificial
network.

Fig. 3. ANNs interconnected with the other cells
4.2 Evolutionary Programming of the ANNs
Wilfried and István used a Java framework called “Frevo” to
evolve the system and study its behaviour with each
generation. The software offers the option of clubbing
different representations and optimization methods for the
problem at hand. In Frevo, each problem has a generic
interface to 1 or many instances of the representation, the
optimizer evolves the control algorithm that was formed.
Fig. 4. Algorithm used for EA
5. Experiments Conducted
The easiest patterns to replicate using cellular automata are
flags, and Figure 2 represents how the Hungarian flag when
fed to the developed model behaves at each generation. The
complexity of the Hungarian and the French flag are similar
as they both are made up of 3 colours and are placed one
after the other
Fig 5. Hungarian flag through generations
It is often observed in evolutionary algorithms that the
cellular automata after around a hundred generations reach
a maximum accuracy state therefore improvements rarely
happen after reaching the local cost minima.
Fig 6. Austrian flag through generations
When subjected to simpler pictures, such as the Austrian
flag, which is made up of 2 colours, a perfect reproduction of
the image is achieved after only 90 generations (Refer to
Figure 6). But when objected to complex paintings such as
the Mona Lisa, the best possible reproduction of the
downsized image, with the dimensions 20x29, wasobtained
after 500 generations. The result had properregeneration of
the background however, the subject, Mona Lisa couldn’t be
processed by the system (Refer to Figure 7). The reason for
this can be the distance between the cells at the edges and
the cells in the middle of the system, as the information
propagates from the edges to the centre of the system.
Fig. 7. CA attempting to recreate Mona Lisa

6. Conclusions
Although the model developed by Wilfried and István might
not be able to generate a constructive solution, the
experiment carried out proves that integrating Artificial
Neural Networks into a Cellular automaton is one possible
approach to designing a Self-organizing System.
The reason the model is unable to come to a feasiblesolution
would be the poor fitness functionused whiledevelopingthe
model. The model for each solution was comparing the
image generated with the reference image, pixel forpixel. As
a result, a solution quite similar to the reference image was
not considered by the system, whereas we humans perceive
a similar pattern as being closer to the reference than an
image in which only parts of the reference image was
reproduced.
The model can be improved by adding more connections
between the neurons in order to reduce the time required to
propagate data from the corner cells to the center and
rewriting the fitness function such that the fitness is based
on the type of emerging structure instead of comparing the
solution pixel-to-pixel with the reference image
REFERENCES
[1] W. R. Ashby, An introduction to cybernetics, Chapman &
Hall Ltd, 1961.
[2] W. Banzhaf, Self-organizing systems., Encyclopedia of
complexity and systems science 14 (2009) 589.
[3] P. J. Bentley, S. Kumar, et al., Three ways to grow designs:
A comparison of embryogenies for an evolutionary design
problem., in: GECCO, Vol. 99, 1999, pp. 35–43.
[4] W. Elmenreich, I. Fehérvári, Evolving selforganizing
cellular automata based on neural network genotypes, in:
International Workshop on Self-Organizing Systems,
Springer, 2011, pp. 16–25.
[5] W. Elmenreich, G. Friedrich,Howtodesignselforganizing
systems, 2009.
[6] I. Fehérvári, W. Elmenreich, Evolutionary methods in
self-organizing system design., in: GEM, 2009, pp. 10–15.
[7] T. Ilango, S. Murugavalli, Advantageofusingevolutionary
computation algorithm in software effort estimation,
International Journal ofApplied EngineeringResearch9(24)
(2014) 30167– 30178.
[8] E. M. Izhikevich, J. H. Conway, A. Seth, Game of life,
Scholarpedia 10 (6) (2015) 1816.
[9] C. Prehofer, C. Bettstetter, Self-organization in
communication networks: principles and design paradigms,
IEEE Communications Magazine 43 (2) (2005) 78–85.
[10] G. D. M. Serugendo, M.-P. Gleizes, A. Karageorgos, Self-
organising systems, in: Self-organising Software, Springer,
2011, pp. 7–32.
[11] A. WR, Principles of the self-organizing dynamic
system., The Journal of General Psychology 37 (2) (1947)
125–128.

A Study on Evolving Self-organizing Cellular Automata based on Neural Network Genotypes

More Related Content

Similar to A Study on Evolving Self-organizing Cellular Automata based on Neural Network Genotypes (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Study on Evolving Self-organizing Cellular Automata based on Neural Network Genotypes