See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221471167

An interactive visualization of Genetic Algorithm on 2-D graph

Conference Paper in International Journal of Software Science and Computational Intelligence · August 2011
DOI: 10.1109/COGINF.2011.6016133 · Source: DBLP

CITATIONS READS
0 1,375

4 authors:

Humera Farooq Nordin Zakaria

Bahria University Karachi Campus Malaysia
22 PUBLICATIONS 127 CITATIONS 105 PUBLICATIONS 741 CITATIONS

SEE PROFILE SEE PROFILE

Mohd Fadzil Hassan Suziah Sulaiman

Universiti Teknologi PETRONAS Universiti Teknologi PETRONAS, Bandar Seri Iskandar, Perak, Malaysia
222 PUBLICATIONS 1,322 CITATIONS 187 PUBLICATIONS 1,320 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Humera Farooq on 12 November 2016.

The user has requested enhancement of the downloaded file.

 An Interactive Visualization of Genetic Algorithm
on 2-D Graph
Humera Farooq M Nordin Zakaria, Mohd Fadzil Hassan and
Computer Science Department Suziah Sulaiman
Future University Computer and Information Sciences Department
Khartoum, Sudan Universiti Teknologi PETRONAS
Email: [email protected] Tronoh, Malaysia
Email: [email protected]

Abstract—The visualization of search space makes it easy to techniques to explore the internal process of the GA, termed
understand the behavior of the Genetic Algorithms (GAs). We as the Interactive Genetic Algorithm (IGA).
proposed a novel way for representation of multidimensional IGA is a technique to design an environment with user
search space of the GA on 2-D graph. This visualization is carried
out based on the gene values of the current generation, and interaction to evaluate solutions. The visualization of the
user interruption is only required after several generations. The search space not only demonstrates the process of searching
main contribution of this research is to propose an approach with GA towards a fitted solution, but it also interacts with the
to visualize the GA search data and to improve the searching search space to help GA to search for the best solution in less
process of the GA with user’s intention in different generations. time. The Interactive Evolutionary Algorithm (IES) was first
Besides the selection of best individual or parents for the next
generation, interference of user is required to propose a new demonstrated by Dawkins [4], when he created a visualize
individual in the search space. Hence, the active user intervention tool to model an artwork called bimorphs. A detail survey
leads to a faster searching, resulting in less user fatigue. The on using Interactive Evolutionary Computation may be found
experiments were carried out by evolving the parameters to in [5] in which, Takagi divided the interactive evolution by
derive the rules for a Parametric L-System. These rules are then two definitions. The first one is a narrow definition according
used to model the growth process of branching structures in 3-D
space. The experiments were conducted to evaluate the ability to which human evaluation is used as the fitness value for
of the proposed approach to converge to optimized solution as an optimized solution , and the second is a broad definition
compared to the Simple Genetic Algorithm (SGA). according to which the optimized solution is obtained from a
Index Terms—Evolutionary Algorithm, Scientific visualization, human-machine interface.
Parametric L-System, Interactive Genetic Algorithm, 2D Graph- In the past, numerous works were done in the field of IGA.
ics
Cartoon style faces [6] were created on coffee mugs by using
the Neural Network with IGA. They used the Neural Network
I. I NTRODUCTION
(NN) to approximate the aesthetic intentions of the user. With
GA is a dynamic random searching algorithm used to solve the learning mechanism of NN, they inferred the desirable
the optimization problems. It was invented by John Holland in solution by the user, whereas IGA was used for evolving
1960’s [1] to explain the adaptive process of natural systems the next generation. In this way, they save the user from the
[2]. GA is based on the biological methods to produce the tiresome work of selecting and assigning fitness to the best
solutions (next population). Being a successful meta-heuristic design. In [7], IGA was used for aircraft design. A large search
technique, this algorithm is widely used for solving many space with multidimensional data was needed for the aircraft
optimization problems. GAs are stochastic search methods that design; therefore interaction in every generation becomes not
are successfully applied widely in the fields of engineering, only infeasible but also impossible. For that reason, they
medicine and machine learning problems. The problem is introduce interaction with the GA search space after a certain
represented in the form of a chromosome, which is converted number of generations, and an objective function was used
into genotype form. the performance of GA depends on many to calculate the fitness for each solution. In their application
factors including a good fitness function, population size and for representing multidimensional data, different graphs were
choice of GA operators (crossover and mutation). used. A fashion design application was developed by [8] in
In computer graphics, visualization plays an important part which the user’s intention is used to evolve the next generation.
in helping the users to understand the problems graphically The fitness for the fitter solution was also assigned by the user.
[3]. Visualization technology raises the level of understanding Applications on GA have extended from computer science to
multidimensional or high resolution data. Another advantage the other fields, i.e. engineering [7], physics, social science,
of using this technology is to reduce the work period. Visual- economics, molecular biology and biomedical engineering [9].
ization does not mean only to view the graphical pictures, but IGA is also used to graphically represent the evolution or
it also includes analyzing and interpreting the data. Over the modeling of plants. The evolved solutions are represented as
past few years scientists have taken advantage of visualization models onto the computer screen and the user evaluates and
selects the parents for the next generation [10], [11]. In [12] solutions in the search space, these visualization techniques
an interactive system is developed, which involves the creation are unable to determine the searching behavior of GA during
of procedural 3-D models of trees using IGA. The user adjusts the evolution process.
parameters to draw a tree. If the user misses some parameters We proposed a new way for interactive visualization of
adjustment, then it can be done by GA. The user can also give the GA search space data. Each generation of the search
ratings (evaluate) to trees for deciding the parents for the next space is displayed as a 2-D graph on the canvas, gives a
generation, but it is not forced to do so in every generation. clear picture of evolving gene values at different locations.
The main focus of their work is not on creating a virtual scene; The user participation is to propose a new individual instead
instead, it’s on giving an idea of how to adjust the parameters of the evaluating and assigning fitness to existing solutions.
interactively. This interaction depends on the convergence of GA present
For generating the plant like structure, GA is also used to as fitness versus generation graph in a separate graph. The
optimized the symbols or parameters for the L-System [13]. proposed individual becomes a part of the search space in the
The L-System uses a rewriting mechanisms such as Chomsky next generation. The output of IGA (fitted individuals) is used
grammar to create the models of the natural environment, to create the rules for generating branching structures using
animations and virtual scenes. Two main parts of the L-System Parametric L-System.
are the axiom and the rules. These rules are recursive by
III. P ROPOSED A PPROACH
nature and are used to model complex plant and tree structures.
To set and arrange these rules by hand is quite a difficult The objective of this study is to develop an interactive
and time consuming job. Therefore, scientists and researchers visualization of multidimensional data (n-D search space) of
introduced a way of using GA for optimized (best answer) and GA on 2D space. For achieving this goal, a tool named
random solutions to generate these branching structures with VIGA-2D (Visualization of Genetic Algorithm on 2-D Graph)
the L-System. was designed and developed. The purpose of developing this
A further extension of the L-System is known as Parametric visualization tool is to make a clear understanding of the
L-System. In the Parametric L-System, the associated param- searching process of GA in every generation and to make it
eters are used with symbols. These parameters are responsible feasible in a way that the user can actively participate during
for giving different angle values to generate trees onto the searching. For interaction, any generation may be interrupted
canvas. In the proposed work, we used an interactive approach by the user. Prior to becoming a part of the search space,
of GA to evolve parameters in order to derive the parametric the fitness of the new individual is compared with the average
rules for modeling the tree structures. The precipitation of fitness of the current generation. Thus, the higher the fitness of
the user is after several generations. The purpose of these the proposed individual, the more chance it has to be select in
interactions is to propose a new solution into search space. the next generation. This interaction also affects the population
size; each successful interaction will increase the population
II. P ROBLEM D EFINITION size. The complete procedure of the proposed approach has
In conventional GA, the search process remains hidden from been outlined in Fig 1. Listed below are the some advantages
the human eye. Only the optimized solution can be viewed of the proposed work.
on the screen. The performance of the algorithm is typically • The proposed interactive approach works with a human-
monitored by a fitness convergence (a graph with fitness versus machine interaction.
generation) rate. Using this graph, the user can only examine • The visualization is based on the gene values in each
the performance of algorithm over each generation. Thus to generation.
analyze these solutions and to change them is quite difficult. • Human interaction is not involved with the selection or
Also, there is less chance to understand the GA convergence or rating the solutions; instead the user proposes a new
the change in the parameters and functions during this search individual in the search space.
process [14]. Moreover, this technique also has not given • There is no need to keep the generation numbers smaller
satisfactory solutions when applied to artistic and aesthetic because the user interaction is not necessary in every
problems. generation.
On the other hand, IGA is found to be a flexible and user • Any generation may be interrupted by the user.
friendly technique for solving artistically and aesthetically • The working mechanism of the GA operators (crossover
relevant problems. The GA visualization gives a suitable and mutation) and selection of individuals work as a
solution for the problem in which the inference of the user hidden process.
is necessary to have an opinion for evaluation and selection of • The GA search process remain continues until the user
solutions [12]. These applications are mostly used for model terminates it.
representations of solutions [15][10][16]. The main problem of The problem is defined as a user input to the VIGA-2D.
traditional IGA is the continuous interaction of the user which It also involves the input of all necessary parameters from
cause fatigue to user and gives unwanted solutions. Moreover, the user which are necessary to initiate the VIGA-2D process.
these approaches require a small number of generations which Then, VIGA-2D is evolved until the user’s perception. The
reduces the user fatigue. Due to the model representation of user may generate the branching structures in any generation
 Start Where parameter is a particular genes value, values is
the range of gene values, m = (0...populationsize) and l =
Next Visualization
Encoding Chromosome
Generation as 2-D graph (0....chromosomelength).
Fitness Function
After retrieving a particular gene value, the next step is to
Selection Method Evaluate

GA operators
genes values
for each
calculate the location of the y and x-axis. The calculation of
generation the y-axis is based on the Vunit as shown in Equation 3.
Add proposed genes
values into search
space Satisfied
N solution

N Y
yaxis = Vlength − (parameter − mvalues ) ∗ Vunit (3)
Increase Y User
population size
interaction
Visualization Where Vlength is the total vertical length of the graph,
Stop of branching
structures parameter is the gene values to display, mvalue is the
minimum range of the gene values and Vunit is the pixel ratio
Fig. 1. An overview of proposed approach along the vertical view. The location of the x-axis is based on
the calculation of the fixed interval as shown in Equation 4.

xaxis = xaxis + interval (4)

and may proceed to the next generation again. With the mul-
tiple interventions of the user in the search space, VIGA-2D Where interval is a fixed value used to keep all data
converges to the optimal solution or the solution according to displays according to the specific gene locations. In this way,
the human perception in fewer generations. Derived parametric each calculated y and x-axis is responsible to display a single
rules are used to model the branching structure onto a 3-D value.
space. The following section describes in detail the technique The last step is involved with the visualization of data onto
used to bring multidimensional data on to 2-Dimensional. the graph according to the x and y-axis. For visualization
A. Representation of the search space on a 2-D Graph of a clear picture of the GA behavior for searching for the
solutions in a 2-D space, the fitness of the gene values are
The visualization of the proposed approach is based on
displayed in blue level colors having a different color depth
the representation of gene values on a 2-D graph. In the
with different sizes. Furthermore, the gene values with the best
past, the attention of most of the researchers was based on
fitness are displayed with the higher fitness having a large
the visualization of solutions of GA in 2 or 3-Dimension
size as compared to the gene values with the worse fitness
[17][18][16] , in which the distribution of gene values and the
values which are displayed in light gray color with a small
searching process of GA remains hidden from the user’s eye.
size. The user selects points (new gene values) according to
For the proposed approach, GA visualization is represented as
the distribution of the gene values with higher fitness for a
a 2-D graph in such a way that all the gene values of each
better proposed solution. The geometrical shape ”ellipse” is
generation are on one graph. The designed graphical interface
used to display gene values onto the graph. To distinguish
is shown in Fig 2. The display of data involves three steps:
between the proposed gene values and the search space gene
The first step is to calculate the ratio to set the data within
values, the proposed gene values are displayed in red color.
a vertical and horizontal range. According to the division
of the graph, for the vertical view, the data depends on the B. User’s Intervention
maximum range of gene values. The ratio for the vertical range
User interaction is the main part of the proposed approach.
is calculated according to the Equation 1. Where vunit is the
Most of the existing IGA applications were based on user
pixel value range along the vertical view, Vlength is the total
interaction in each generation, which creates a tedious en-
vertical length of the graph, GAP is a value used to keep
vironment for the user. To overcome this problem, several
all the data display inside the graph and mrange is maximum
techniques were introduced in the past, i.e. approximation
range of gene values.
of fitness values using a neural network [19] or to assign a
discrete fitness value for the evaluation of solutions [20]. How-
Vunit = Vlength − GAP/mrange (1)
ever, these existing techniques worked with a fixed range of
For the horizontal view, the division of the graph depends parameters and the fitness values. It was also found that these
on the chromosome length. The graph is equally divided into existing applications were based on the subjective selection of
the fixed length intervals. Hence, each interval represents each the parents for the next generation.
gene location. In the proposed work, user is not involved with the selec-
The second step is to retrieve a particular gene value from tion of the existing solutions. The user may select the gene
a chromosome in each generation. Gene values were retrieved values on graph after proceeding to many generations. For
from every location of the chromosome in each generation that reason, the searching process is divided into two parts:
according to the following Equation 2: (1) there is no interaction in the current generation by the
user; hence, the VIGA-2D does the convergence normally.
parameter = valuesm,l (2) (2) The user interacts the current generation and proposed a
new solution. Then, VIGA-2D first calculates the fitness of
the new solution to ensure that the proposed solution is fit
enough. For that reason the fitness of the proposed solution is
compared with the average fitness of the current generation. If
the proposed solution’s fitness is higher than the average fitness
of the current generation, then the population size increases by
one and make it is made as a part of the search space otherwise
discard it. After that, crossover, mutation and selection are
performed on a new search space.
The designed graphical interface of VIGA-2D helps the
user to decide the generation for interaction. For example two
different colors, i.e. blue and green, are used to demonstrate
the current fitness. If the green color is highlighted, it means
Fig. 2. Graphical Interface for VIGA-2D
that the current generation has a less average fitness as
compared to the previous one, and the user may interrupt
the searching process to improve the overall fitness. If the
highlighted color is blue, it means that the GA searching is v
u l
already towards a fitted solution. Each interaction is based on uX
the upper and lower bounds of particular gene values. f =t (Gparami − Oparami )2 (6)
i=1
Since the user interacts directly with the graph, so this
interaction is based on the pixel values. These pixel values Where f is the fitness value, Gparam is the GA solution and
are converted into data values (gene values) according to the Oparam is the target solution and i = (0...l) is the length of the
calculation of the x and y-axis. The y-axis is involved in chromosome. The difference of both arrays is the fitness value
retrieving data according to the upper and lower bounds of of the current evolving individual. This evaluation function
each gene value. Equation 5 is used to convert the pixel values analyze the fitness measures for each individual indicating its
into data values. suitability to be selected.
Different selection methods are used to check the perfor-
mance of GA with the proposed approach. Different mutation
parameter = (vlength − ypixel )/(vunit + mvalue ) (5) and crossover rates are also applied to check the efficiency of
the proposed algorithm. However, in our presented results, we
Where parameter is the gene value to display, Vlength is have used 1-point crossover and 1-point mutation. The details
the total vertical length of the graph, ypixel is the y − axis for using all these GA operators and the selection method
position on the graph, vunit is the pixel ratio along the vertical in our work can be found in [21]. VIGA-2D evolved until
view and mvalue is the minimum range of gene values. it found the optimal solution either according to the target
The x − axis is involved to specify the gene location for solution or according to the user perception for the growth
each gene value. The length of the chromosome may vary structure. During the searching process, the user may model
according to the input of the user. The maximum chromosome the branching structure at any generation. For modeling the
length is 30. As discussed above x−axis is divided into equal structures, rules are derived from the fitter solution of the
intervals according to maximum length of the chromosome. current generation. For deriving the rule, the color values in
These intervals are used to convert pixel values into data values Axiom are not evolved by the VIGA-2D. The Axiom received
for x − axis. two parameters with a symbol A i.e.
ω = (r, g, b)A(x, w)
C. VIGA-2D for Parametric L-System
Where (r, g, b) are color values for Red, Green and Blue,
Parametric L-System consists of two main parts, i.e. sym- and A is used for recursive purpose. The x shows the paramet-
bols, which are responsible for directions; and parameters, ric value for the line symbols F and w shows the parametric
which are responsible for the angles of the turtle interpretation. value for the width.
Turtle interpretation is used to visualize L-System rules. A For the L-System, rules can be vary in numbers, depends
turtle is an object which is moved on a Graph (F) and rotates on the complexity of generated structure. We developed one
in three directions, i.e. (U axis, H axis and L axis). These turtle rule from the output of V IGA − 2D. Construction of the rule
graphics are used to build a geometrical interpretation of L- is as follows:
System strings. In the proposed work, the direction symbols
A(x, w)− >!(w)F (x)[−(θ1 )A(r1 , wr1 )]F (x1 )[+(θ2 )
are given by the user in the form of an L-String. Then,
A(r2 , wr2 )]F (x2 )[+(θ3 )A(r3 , wr3 )]
with the help of this L-String, a random chromosome (Target
...F (xn )[−(θn )A(rn , wrn )]
solution) is constructed. Equation 6 shows the fitness function
to calculate the difference between the target and the GA According to these rules following are the assumptions
solution. and limitations used to make rules for generating branching
structures. to the human perception after generating the branching struc-
• For assigning parameter to the symbol () are used tures in different generations. The evaluation and effectiveness
i.e.!(2.53) shows that the width of branches is 2.53. of the proposed algorithm is compared with SGA with the
• The predecessor may also receive variable using () i.e. same parameters. This comparison is based on the convergence
A(x, w). This variable is used in the successor as param- test, i.e. the average and best fitness and drawn structures by
eter. both algorithms. Finally, we did a statistical analysis between
• In production rule, symbol A behave recursively for n our proposed approach and the traditional IGA. The following
number of iterations, where value of n is the number of were the rates and operators used for this experimental result:
iterations. • Roulette wheel selection was used.
• The number of iteration shows the complexity of the • Random one point crossover was used.
generated structure. • One-point mutation was performed with the 0.05 muta-
• For modeling the branching structures two modules F tion rate.
and A was used. F was responsible for producing the • Initial population size was = 10.
line and A was used as interpretation point. • Iteration number for generating structures = 02.
• The chromosome length for parameters depends on the • Input by user (L-String) = !F [−F [+F ]F ][−F ].
L-String given by the user as an input.
For example if the L-String given by user is =!F [+]F [−], Fig 3(a) shows the structure drawn by the L-String and the
angle for each branch θ is 22 and initial parameter are: initial parameters, ( the width and height of the trunk) were
37, 4, 28. Then after initial structure the L-String becomes: given by the user. The VIGA-2D process is initiated with a
!F [+!F [+]F [−]]F [−!F [+]F [−]] and the length of random random chromosome. The length of the chromosome was 12
chromosome will be 21 , based on the total number of symbols including the 2 parameters for the symbol A for each direction.
in L-String. Fig 3(b) shows the drawn structure for next iteration with
After evolving the VIGA-2D for n-number of generations, L-String. The branching structures generated with SGA and
following is the derived rule with integration of an L-String VIGA-2D show a clear difference in the parametric values. For
and the best solution obtained from VIGA-2D: example, for generation 52, the rule generated with VIGA-2D
has the following form:
ω = (r, g, b)A(19, 8) Axiom: (0.24,0.6,0)A(29,4)
#A(x,w)->!(w)F(x)[-(22)$A(x*0.84,w*0.56)
$A(x,w)->!(w)F(x)[+(22)$A(x*0.81,w*0.81) F(x*0.81)[+(22)$A(x*0.5F,w*0.56)
!(w*0.81)F(x*0.81)[+(22)[-(22)$ (x*0.43)]F(x*0.65)][-(22)$A(x*0.81
A(x*0.81,w*0.51)]F(x*0.35)[-(22) ,w*0.79)F(x*0.31)]
$A(x*0.61,w*0.21)]]F(x*0.81)$A(x And the rule generated with SGA for same generation is as
*0.53,w*0.65)!(w*0.31)F(x*0.69)[ follows:
+(22)$A(x*0.72,w*0.88)]F(x*0.76)
[-(22)$A(x*0.51,w*0.45)]] Axiom: (0.24,0.6,0)A(8,4)
#A(x,w)->!(w)F(x)[-(22)$A(x*0.35,w*0.79)F
(x*0.31)[+(22)$A(x*0.37,w*0.81)
IV. E XPERIMENTS F(x*0.62)]F(x*0.81)][-(22)$A
GA is a search technique used in computing field to find (x*0.83,w*0.88)F(x*0.89)]
exact or approximated solutions to optimization and search
problems [22]. Finding of an optimized solution of GA de- The main difference in both rules was noted for the param-
pends a lot on the nature of the problem. If the solution is eter values for symbol A in the axiom,which are the first two
integer or string based, then to find the optimized solution gene values in the chromosome. This difference in parameters
is easier as compare to the problem in which the parameters brings a difference in branching structures as shown in Fig 3
have real values. In the present research, the problem given c and d.
to GA was to optimize the parameters for L-System rules
which were real values. Therefore, most of the chances were A. Convergence Test
to get the nearest optimized solution as compared to getting The user evolved VIGA-2D for 85 generations. Termination
the optimized solution. of the process was according to the human perception. SGA
In order to show the effectiveness of the proposed approach, was also evolved until 85 generations. Fig 3 shows the
we invited 5 subjects to run VIGA-2D with different para- drawn structures with same parameters with and without sev-
metric values and L-Strings. For GA, different crossover and eral interactions. The structures were generated at generation
mutation rates were tested. Among these experiments, one 19, 52, 69and85. In total, 14 times the VIGA-2D process was
result has been selected to present here. The objective of the interrupted by the user to propose a new solution. Among
presented result is to terminate the searching process according them, 2 proposed individuals were discarded by VIGA-2D.
 TABLE I
A STATISTICAL DETAIL FOR DEMO RESULT

Parameters VIGA-2D SGA

Population size
before Interaction 10 10
Total Interactions 14 N/A
Accepted 12 N/A
Interactions
Discarded 02 N/A
Interactions
Population size
(a) Best Fitness Graph after interaction 22 10
Average Fitness 0.583039 0.61818
Best Fitness 0.386652 0.58923

TABLE II
L IST OF ACCEPTED AND DISCARDED SOLUTIONS AT DIFFERENT
GENERATIONS WITH FITNESS VALUES

Modeled
at Generation 19 52 69 85
Interaction at 8,16 28,32, 57,61, 72,77,
Generation 42,49 66,69 82,85
Best Fitness
(b) Average Fitness Graph for VIGA-2D 0.32311 0.47497 0.37536 0.38665
Best Fitness
Fig. 4. Fitness/Generation Graph for SGA 0.41988 0.55434 0.36687 0.58923
Discarded at
Generation 16 Null 61 Null

For VIGA-2D, the population size at the termination point was

22. The first structure was generated at generation 19 for both
algorithms. The best fitness for VIGA-2D was 0.32311 and gene values at some locations were unnecessary increasing
for SGA was 0.41988 at this generation. It was noted that the in generations which was bringing the GA quite far from its
structure generated at generation 69 with both algorithms had a target solutions. To overcome this problem, the interaction
minor difference between fitness as shown in Table II. The user would be an immediate action to bring the gene values to
continued with the next generation and generated a structure at the desired frequency in the next generations.
generation 85, which showed a difference in fitness for both While doing the experiments with users for generating
algorithms. Table I shows the entire parameters list and the the branching structures, the understanding of the users with
difference of average and best fitness for both algorithms. the proposed approach was closely monitored. It was noted
The Fitness / generation graph in Fig 4 shows the con- that there was no consistency in doing the interactions, for
vergence difference for SGA and VIGA-2D for this experi- example, some users did all the interactions in the initial
mental result. According to Fitness / Generation graph, the generations while some users evolved many generations with-
convergence of SGA and VIGA-2D was noted at the same out any interactions. Although the overall performance and
fitness value in the initial generations. It was also noted that decision for human interaction also depends on the fitness /
for VIGA-2D the average fitness was higher as compared to generation graph, but during the interactions the users were
SGA for a few generations. However in the last 20 generations more interested to viewing the output as compared to moni-
a prominent difference was noted between fitness values of toring the convergence rate. Moreover, while doing different
both algorithms. experiments, it was observed that some users were more
interested in modeling the branching structures as compared
B. Observations and User Analysis to understanding the VIGA-2D environment.
Visualization of gene values as a 2-D graph is proposed to The decision to go for the interactions was a critical point
evaluate the frequency of genes at particular locations and to in the proposed algorithm. A continuous observation had been
do various interactions. The evaluation of results with different done on the users to gauge their understanding for the next
users was helpful to analyze the performance of VIGA-2D. interactions. These observations were as follows:
The proposed graphical interface gives a clear picture of the • The user interacts when there were fewer numbers of
multidimensional search space of the GA and makes it easy for gene values at a particular place,
a user to monitor the frequency of gene values at a particular • the genes values’ color intensity was low at any gene
place. During the search process, it was noticed that, proposing location,
new gene values at different locations helps to fill up the gap • the user feels that a group of gene values were at the
between the evolution processes. It was also noticed that the same location for many generations,
 (a) Generated with L- (b) Next iteration (c) SGA Generation (d) VIGA-2D Gener- (e) SGA Generation (f) VIGA-2D Genera-
String level with L-String 19 ation 19 52 tion 52

(g) SGA Generation (h) VIGA-2D Gener- (i) SGA Generation (j) VIGA-2D Genera-
69 ation 69 85 tion 85

Fig. 3. Corresponding Branching structure with SGA and VIGA-2D.

• the proposed color is green and it may produce unnecessary noise and distortion in the search
• when the convergence of the fitness according to the space. But due to using the objective function the GA may
fitness graph was not converging towards the fitter value. solve this problem in a few generations or it could be solved
The termination criteria were according to the user perception. more quickly by doing more interactions.
It was observed that there were several assumptions and ob-
servations taken by different users to terminate the procedure. C. Comparison with Traditional IGA
For example: Table III shows the difference between the proposed ap-
• According to the user, the distribution of gene values was proach and TIGA. For the traditional IGA applications, the
equally distributed. applications were based on the constant number of generations
• Most of the genes value’s color intensity was high (Dark [10][11] in which the interaction of the user is necessary
Blue Color), means the majority of the search space had in each generation for evaluating and assigning the fitness
fitter solutions. to the solutions. Moreover, in these applications, the user’s
• There was not any prominent difference between the perception is used to evaluate the fitness of solutions in a
distributions of gene values even with the interactions. population. These existing IGA applications gave a suitable
• There was no prominent difference between the generated solution for the problem in which the inference of the user is
structures in last few generations. necessary to have an opinion for evaluation and selection of
• Convergence of the fitness according to the fitness graph solutions [12]. The method used in these applications helps
was not converging towards a fitter value. the user to draw or select the individuals according to a visual
The color intensity also played an important role for the picture of that object in his mind. In this way, in TIGA, most of
decision of interactions. The user observed the variation of the applications are completely dependent on the user. On the
color and size of the ellipse with more interest as compared other hand, in human- machine interactions, the applications
to the fitness convergence. are based on the user’s preferences and selections or it may
The impact of variation in population size was also ana- use some classifiers to approximate fitness values.
lyzed. It was observed that the increment in population size The proposed algorithm is different from these applications
also impacted on the performance of the searching process. in that the interaction of the user is not necessary in each
With each interaction, the search space becomes wider. In generation. The fitness of the solutions is calculated with the
most of the experiments, the resultant population size was 50 objective function and there is no need to evaluate the solutions
percent bigger than the initial size. This feature gives a facility for the next generation.
to use a wider search space interactively.
On the other hand, our proposed methodology gives a V. CONCLUSIONS
chance to GA to make a decision to absorb the proposed values Research efforts made to understand the research presented
as a part of the evolution or discard them. In this way, the in this paper focus on exploring the technique to visualize
user performance could be easily monitored with the effect the multidimensional data of the GA search space. Based on
on population size. It was also observed that proposing gene the proposed method, a tool (VIGA-2D) has been developed in
values having worse fitness or allocating null values, directly which each generation is displayed visually. This visualization
impacted on the performance of the algorithm. For instance , gives a clear picture of evolving gene values at different
 TABLE III
C OMPARISON BETWEEN PROPOSED APPROACH AND TIGA [9] G. Kkai, R. Vnyi, and Z. Tth, “Parametric l-system description of the
retina with combined evolutionary operators,” in Genetic and Evolution-
ary Computation Conference, GECCO, 1999.
Parameters VIGA-2D TIGA [10] H.-J. Min and S.-B. Cho, “Creative 3d designs using interactive genetic
Population size Variant de- algorithm with structured directed graph,” pp. 391–400, 2004.
pends on [11] R. Curry, “On the evolution of parametric l-systems,” 2000.
how many Constant [12] C. Clare Bates and H. M. Raymond, “Gentree: An interactive genetic al-
interaction gorithms system for designing 3d polygonal tree models,” in Genetic and
Generations Based on Evolutionary Computation GECCO 2003, vol. 2724/2003. Springer
user Constant Berlin / Heidelberg, 2003, p. 216.
Perception [13] P. Prusinkiewicz and A. Lindenmayer, The algorithmic beauty of plants.
Interaction Proposed Select from Springer-Verlag New York, Inc., 1990.
new [14] H. NISHINO, H. TAKAGI, and K. UTSUMIYA, “A 3d modeler for
Individual existing individuals aiding creative work using interactive evolutionary computation,” IEICE
Termination Based on Pre-defined Transactions on Information and Systems, vol. 85, pp. 1473–1483, 2002.
Criteria User Per- Criteria [15] S.-B. Cho, “Towards creative evolutionary systems with interactive
ception genetic algorithm,” Applied Intelligence, vol. 16, no. 2, pp. 129–138,
2002.
[16] H. Nishino, H. Takagi, C. Sung-Bae, and K. Utsumiya, “A 3d model-
ing system for creative design,” in Proceedings of 15th International
Conference on Information Networking, 2001.
locations. The fitness is calculated using the objective function [17] U. Aybars, “Path planning on a cuboid using genetic algorithms,”
and the GA selection method is responsible to select the fitted Information Science, vol. 178, pp. 3275–3287, 2008.
[18] I. Shin-Ichi, M. Yasue, M. Hiroko Nakamura, S. Takafumi, and F. Mi-
solution for the next generation. noru, “A visualization of genetic algorithm using the pseudo-color,” in
During searching process, a user may interact to propose a Neural Information Processing: 14th International Conference, ICONIP
new individual. The proposed approach does not depend on 2007, Kitakyushu, Japan., 2008.
[19] B. John, A. Peter, and L. Laura, “Neural network fitness functions for
continuous interaction of user. The interaction may be done in a musical iga,” in International Symposium on Intelligent Industrial
any generation and any number of times until a satisfactory Automation and Soft Computing, 1996.
solution is obtained. The proposed individual become a part [20] M. Ohsaki and H. Takagi, “Improvement of presenting interface by
predicting the evaluation order to reduce the burden of human interactive
of search space in next generation. The developed approach ec operators,” in IEEE International Conference on Systems, Man, and
has been used to evolve the parameters for deriving rules Cybernetics., 1998.
for Parametric L-System. These rules are then generate the [21] F. Humera, Z. M. Nordin, H. Mohd. Fadzil, and S. Suziah, An Approach
to Derive Parametric L-System Using Genetic Algorithm. Springer-
branching structures on a 3-D canvas. The performance of Verlag, 2009.
VIGA-2D had been analyzed with the help of subjective [22] W. Chin, “A visual genetic algorithm tool,” Master’s thesis, Faculty of
analysis. For this purpose, real values were evolved by the Computer Science,University of Malaya, 2006.
proposed approach to derive the rules for Parametric L-
System. The experiments performed with VIGA-2D shows
that it has the ability to perform efficiently as compare to
SGA. Furthermore, experiments show that the performance of
VIGA-2D really depends on the user interactions. Future work
involves analyzing the performance of the proposed work with
other applications.

R EFERENCES
[1] J. H. Holland, Adaptation in natural and artificial systems. MIT Press,
1992.
[2] P. Haroun, “Genetic algorithm and data visualization,” Ph.D. disserta-
tion, 1997.
[3] H. C. Purchase, N. Andrienko, T. Jankun-Kelly, and M. Ward, Theoret-
ical Foundations of Information Visualization. Springer-Verlag, 2008.
[4] R. Dawkins, “The evolution of evolvability,artificial life,” in Interdisci-
plinary Workshop on the Synthesis and Simulation of Living Systems,
C. Langton, Ed. Redwood City,CA: Addison-Wesley, 1989.
[5] H. Takagi, “Interactive evolutionary computation: fusion of the capabil-
ities of ec optimization and human evaluation.” in Proceedings of the
IEEE, vol. 89, 2001, pp. 1275–1296.
[6] Z. Gu, M. Xi Tang, and J. H. Frazer, “Capturing aesthetic intention
during interactive evolution,” Computer-Aided Design, vol. 38, no. 3,
pp. 224–237, 2006.
[7] O. Bandte, “A broad and narrow approach to interactive evolutionary
design–an aircraft design example,” Applied Soft Computing, vol. 9,
no. 1, pp. 448–455, 2009.
[8] R. Jie, G. Dun-Wei, S. Xiao-yan, Y. Jie, and L. Ming, Interactive Genetic
Algorithms with Variational Population Size, D.-S. H. Kang, K.-H. Jo,
H.-H. Lee, Hee-Jun, Bevilacqua, and Vitoantonio, Eds. Springer, 2009,
vol. 5755.

View publication stats