Dynamic Radius Species Conserving Genetic Algorithm for Test Generation for Structural Testing

International Journal of Software Engineering & Applications (IJSEA), Vol.8, No.2, March 2017
DOI: 10.5121/ijsea.2017.8208 121
DYNAMIC-RADIUS SPECIES-CONSERVING GENETIC
ALGORITHM FOR TEST GENERATION FOR
STRUCTURAL TESTING
Michael Scott Brown and Michael J. Pelosi
Information Technology and Serveries Department, University of Maryland University
College, Adelphi, Maryland, United States
ABSTRACT
Software testing is a critical and labor-intensive activity in software engineering. Much research has been
done to help automate test case generation. This research proposes a new approach to structural test case
generation. It uses a specialized genetic algorithm called Dynamic-radius Species-conserving Genetic
Algorithm (DSGA) to find a structurally complete set of test cases for the Triangle Classification algorithm.
DSGA is a Niche Genetic Algorithm (NGA) that uses a short-term memory structure to store optima. Each
individual of the NGA represents the inputs for a test case. The fitness function encourages the algorithm to
locate test cases that cover large areas of the structure of the program. A shared fitness encourages the
NGA to locate other areas of the structure. DSGA is a novel approach to structurally complete test case
generation.
KEYWORDS
Niche Genetic Algorithm, Genetic Algorithm, Software Testing, Automated Test Case Generation,
Structural Testing
1. INTRODUCTION
Software testing can be very time consuming and difficult. Some studies show that as much as
50% of software development effort is used to test software [1]. Much of this time is spent
generating test cases that can meet a testing objective. Much research has been conducted to
automate software testing. Research covers generated test data for all types of testing like
structural test cases [2], unit testing [3] and functional testing [4]. Automation of testing allows
for software to be developed quicker and at a lower cost through reduced cost in manual
generation of test cases.
Structural Testing is one of many categories of software test cases. In Structural Testing the goal
is to test all areas of the program based upon the structure of the program. The simplest form of
Structural Testing is Statement Testing. In Statement Testing the goal is to obtain a set of test
cases that execute every statement in the program. A set that executes every statement is said to
be statement complete [5]. In most software not every statement can be executed. Exception
handling is a primary example. In these cases it can be stated that some percentage of statement
coverage has been achieved. In addition to statement coverage, Structural Testing can also be
done for paths, branches, methods, classes and other structural characteristics.
One method often used for software testing automation is the Genetic Algorithm (GA) [6]. GAs
are a category of optimization algorithms that work very well at searching large complex domain

122
spaces. In software testing GAs have been used for test case reduction [7], input test generation
[8] and even fixing software defects [9]. GAs demonstrably perform well in the automation of
software testing.
This paper presents a new GA for the automation of input test generation for structural testing.
This research uses a highly specialized GA called the Dynamic-radius Species-conserving
Genetic Algorithm (DSGA) to generate input data for structural completeness [10]. DSGA for
Structural Tests (DSGA-ST) works in a manner that is similar to how humans generate test cases
for structural testing. It begins by locating some input values and determines what parts of the
structure they are testing. Then it excludes those areas of the domain to search for new inputs that
test other parts of the structure. This approach helps the GA quickly locate input data for test
cases.
This research makes a number of contributions to the area of structural testing field research.
DSGA-ST develops an entire test suite with a single pass running of a GA, while other methods
have to run GAs multiple times for each condition within the target program. Algorithms of this
type generate completely new fitness functions for each condition within the target program.
Also, with DSGA-ST the size of the test suite does not need to be known. Some other algorithms
model the test suite as an individual in the GA or restrict the test suite size to that of the
population size in the GA. DSGA-ST is adaptive in nature and can identify test suites for any
size. This allows DSGA-ST to develop test case suites without the limitation of other algorithms.
2. LITERATURE REVIEW
There is a lot of research on Structural Testing. It is one of the earlier forms of testing and has
shown to be very effective at finding defects [5]. This is a mature area of research, with decades
of cumulative experience, making an exhaustive survey of the area challenging in its own right.
Even when considering using GAs for Structural Testing the number of research papers published
is very large [6].
Therefore, our discussion only surveys a subset of papers on GAs and a review of related papers
on the topic of using GAs for structural testing. This research uses the Triangle Classification
algorithm to obtain results, so we employed the Triangular Classification algorithm as a
benchmark, to provide a cardinal metric of performance gains, which is readily compared.
2.1. Genetic Algorithms
GAs have proven very useful in solving optimization problems. They have been used for decades
to solve problems with large complex domains. Not only have they been used for structural
testing, they have been used for countless other problems like predicting stock prices [11],
optimization of large complex functions [10] and determining how amino acids fold to create
proteins [12].
GAs search through the domain in a similar manner to that via which species evolve to adapt to
their environment. A species of plants or animals have a genetic sequence and different values
for these genes produce individuals with different traits. The species lives in some environment
and in this environment individuals with certain traits do better than other individuals.
Individuals that have traits that do well at solving the problem of existing in the environment are
more likely to reproduce and pass their traits onto their offspring. Over time the species finds a
small set of optimal gene values to cope with the environment [13].
GAs are computer algorithms that model this evolution process that species go through. GAs
have individuals which represent a possible solution to the problem. A set of individuals is a

123
generation. The algorithm uses genetic operations on a generation to produce a new generation,
which is better at solving the problem than the previous generation. Over many generations,
possibly hundreds or thousands, the individuals in the current generation converge to the optimal
set of genes and thus to the optimal value in the domain.
GAs implement individuals which represent a possible solution to the problem that the GA is
attempting to solve [14]. Individuals are represented as strings, most often comprising 1’s and
0’s. Different combinations of these 1’s and 0’s represent different values of the domain. In
problems of functional optimization these would represent values for the input parameters to the
function. But individuals could model much more complex entities, like classification rules [15-
16]. A GA is initialized by randomly generating some number of individuals. This is termed the
first generation.
Accompanying the individual is a fitness function. The fitness function accepts an individual as a
parameter and returns how well the individual solves the problem. The fitness function is used to
determine the best individuals within a generation. Just like in nature the traits of these
individuals are more likely to be passed onto future generations.
The GA will perform selection, crossover and mutation iteratively, with each iteration creating
the next generation. In the first step of the iteration, selection, the GA selects pairs of individuals
for crossover. Individuals with higher fitness are more likely to be selected and individuals can be
used in multiple selection pairs. Eventually each pair of individuals from the selection step will
produce two new individuals. To keep the number of individuals in each generation consistent,
the selection process will produce pairs equal to half of the number of individuals in a generation.
Because individuals with higher fitness are more likely to be selected, their genes are more likely
to be pass onto the next generation.
The next step is crossover. In the crossover step, a position is randomly selected between the
characters in the individual and each individual is split along this position. Switching the tails of
the individuals creates two new individuals. These two new individuals become part of the next
generation. An example of crossover can be seen in Figure 1.
Figure 1. Crossover Example
The final step is mutation. Mutation alters the value of a character based upon a parameter,
mutation rate. When an individual has binary characters, the mutation operator just switches the

124
gene value between 0 and 1. When an individual has genes with multiple values another random
number is produced to determine the new gene value.
GA theory argues that over many generations the individuals in the population will converge to
the correct answer [17]. There are two opposing forces in a GA. Selection and crossover are
exploitive forces that work to narrow the generation into optima. Mutation is an explorative force
that introduces new areas of the domain for consideration. Over many generations the individuals
within the population converge to optima.
2.2. Employing the GA as Structural Test Generator
Because GAs are appropriate for searching, they are well suited for structural test case generation.
A significant number of research papers have been published on using GAs for test case
generation. We have identified more than a 100 works only on the subject of using GAs for
software testing. This review only lists a few relevant ones. We survey only those of specific
relevance.
2.2.1. GADGET
Genetic Algorithm Data Generation Tool, GADGET, [18] uses a coverage table to track the areas
of the program that test cases have been created for. GADGET creates a specialized objective
function for each entry on the coverage table. So if a target condition is if (c >= d), the fitness
function would be:
=
− , ≥
0, ℎ
(1)
GADGET uses a fitness function which is a model of a minimization problem, so this function
would encourage a solution that has c >= d. When test data is found the coverage table is updated
with test case values. This opens the possibilities for coverage within the body of the if statement
to be located.
Michael, McGraw and Schatz [18] research shows results against a number of programs. Table 1
shows the results for the Triangle Classification algorithm as reported in Sofokleous and Andreou
[2]. It produced 9 test cases that have 79.6% coverage over the edges and conditions [2].
A limitation to GADGET is that the GA needs to be run multiple times each time finding a test
case. Different fitness functions are derived for the conditions within the target program. Test
cases for some areas of the target program cannot be obtained until test cases for previous areas
are obtained. This occurs when targets are in the body of conditional statements [2].
Table 1. GADGET results for TC Algorithm as reported in [2]
Test
#
I j k Edges Conditions
1 1680498885 1961702355 -1490056820 1, 2 A1=F; A2= F; A3=T
2 1293470477 1898197634 465181194 1, 3, 4, 7, 10,
13, 14, 15, 16
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=F;
E1=T; F1=F; F2=F;
F3=T
3 -120192928 1041962067 280365949 1, 2 A1=T
4 841354299 -
1802686561
-209782592 1, 2 A1=F; A2=T
5 1056804119 660913846 1617709752 1, 3, 4, 7, 10,
13, 14, 17, 18
A1=F; A2= F; A3=F;
B1=F; C1=F; D1=F;

125
E1=T; F1=F; F2=F;
F3=F
6 719320455 507534636 574028437 1, 3, 4, 7, 10,
13, 14, 17, 18
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=F;
E1=T; F1=F; F2=F;
F3=F
7 743820356 743820356 1826109949 1. 3, 4, 5, 6, 10,
13, 19, 22, 25,
28, 31, 32
A1=F; A2=F; A3=F;
B1=T; C1=F; D1=F;
E1=F; G1=F; H1=F;
I1=T; I2=T
8 999699718 584551117 999699718 1, 3, 4, 7, 8, 9,
13, 19, 22, 25,
26, 27
A1=F; A2=F; A3=F;
B1=F; C1=T; D1=F;
E1=f; G1=F; H1=F;
I1=T; I2=T
9 799340978 1321708382 1321708382 1, 3, 4, 7, 10,
11, 12, 19, 22,
28, 29, 30
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=T;
E1=F; G1=F; H1=F;
I1=F; J1=T; J2=T
2.2.2. Automatic Test Cases Generation System
The Sofokleous and Andreou [2] algorithm comprises of two sub-systems. The Basic Program
Analyzer System (BPAS) analyzes the program, produces a Control Flow Graph and determines
code coverage. The Automatic Test Case Generation System (ATCGS) searches the input space
for an optimal set of test cases for structural coverage [2]. ATCGS comprises of two algorithms:
Batch-Optimistic (BO) and Close-Up (CU).
The BO employs a variant of McCabe’s Cyclomatic Complexity formula to determine the
number of test cases needed for coverage. A complete set of test cases is then modeled as an
individual for the GA. The BO employs the following fitness function:
=
# + "##$ % + #$ &'() *
+ "
2
In this fitness function w1 and w2 are weights. #edgesexecuted is the number of edges executed
by the test cases and #predtrue and #predfalse are the number of predicates evaluated to true and
false respectively. BO stores these test cases in a repository.
The CU algorithm attempts to find test cases for other conditions in the target program not
located by BO. It creates a GA for each condition and where the GA cannot locate a suitable test
case, it marks the target area as unreachable. Table 2 shows the results of the Sofokleous and
Andreou algorithm [2] on the Triangle Classification algorithm. It produced 10 test cases and
provided 100% coverage over the edges and conditions.
Table 2. Automatic Test Cases Generation System results for TC algorithm [2]
Test
#
I J K Edges Conditions
1 93 -1 0 1, 2 A1=F; A2= T
2 -8 -1 -7 1, 2 A1=T
3 5 784 -1 1, 2 A1=F; A2= F; A3=T
4 786 732 1 1, 3, 4, 7, 10, A1=F; A2= F; A3=F;

126
13, 14, 15, 16 B1=F; C1=F; D1=F;
E1=T; F1=F; F2=T
5 476 645 537 1, 3, 4, 7, 10,
13, 14, 17, 18,
A1=F; A2= F; A3=F;
B1=F; C1=F; D1=F;
E1=F; F1=F; F2=T
6 1 1 45 1, 3, 4, 5, 6,
10, 13, 19, 22,
25, 28, 31, 32
A1=F; A2= F; A3=F;
B1=T; C1=F1; D1=F;
E1=F; G1=F; H1=T;
H2=F; I1=F; J1=F
7 1 1 1 1, 3, 4, 5, 6, 8,
9, 11, 12, 19,
20, 21
A1=F; A2=F; A3=F;
B1=T; C1=T; D1=T;
E1=F; G1=T
8 346 1 597 1, 3, 4, 7, 10,
13, 14, 15, 16,
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=T;
E1=F; G1=F; H1=F;
I1=F; J1=T; J2=T
9 5 881 5 1, 3, 4, 5, 6,
10, 13, 19, 22,
25, 28, 31, 32
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=T;
E1=F; G1=F; H1=F;
H2=F; I1=T; I2=F; J1=F
10 1 6 6 1, 3, 4, 7, 10,
11, 12, 19, 22,
25, 28, 29, 30
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=T;
E1=F; G1=F; H1=F;
I1=F; J1=T; J2=T
11 31 4 31 1, 3, 4, 7, 8, 9,
13, 19, 22, 25,
26, 27
A1=F; A2=F; A3=F;
B1=F; C1=T; D1=F;
E1=F; G1=F; H1=F;
I1=T; I2=T
12 15 15 2 1, 3, 4, 5, 6,
10, 13, 19, 22,
23, 24
A1=F; A2=F; A3=F;
B1=T; C1=F; D1=F;
E1=F; G1=F; H1=T;
H2=T
13 4 1 1 1, 3, 4, 7, 10,
11, 12, 19, 22,
25, 28, 31, 32
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=T;
E1=F; G1=F; H1=F;
I1=F; J1=T; J2=F
14 36 42 5 1, 3, 4, 7, 10,
13, 14, 15, 16
A1=F; A2=F; A3=F;
B1=F; C1=F; D1=F;
E1=T; F1=F; F2=F; F3=T
2.3. Triangle Classification Algorithm
The Triangle Classification algorithm is a short algorithm that determines if three lengths would
form a triangle and, if so, would the triangle be equilateral, isosceles or scalene. The algorithm
accepts three parameters, i, j and k, which are the length of the three sides. The algorithm returns
an integer: 1 for a scalene, 2 for an isosceles, 3 for an equilateral and 4 for not a triangle. The
algorithm is often used in structural testing because it is short, but it has many paths and
conditions. Figure 2 shows a control flow graph of the algorithm. Conditional statements are
indicated with a letter. If it is a compound condition, each part of the condition is given a number.
These are shown in the control flow graph in brackets [ ]. For example [A1] indicates the

127
conditional statement A and the first Boolean expression within the conditional statement. These
annotations are used in the results section.
The Triangle Classification algorithm is novel for structural testing for another reason. Assuming
that the input parameters can be negative, a vast majority of the possible test case inputs lead to a
single path in the control flow. No triangle can have a side of negative or 0 length. Condition A of
the control flow checks to see if i, j or k is less than or equal to 0. If so, the three sides cannot
make a triangle and it follows path 2 and terminates. If i, j and k have a sign bit, then a vast
majority of input data follow this, not so interesting, path through the program. Algorithms that
attempt to generate structural test cases need to explore these other areas of the search space.
The Triangle Classification algorithm is a benchmark algorithm for structural testing. In addition
to GADGET [18] and Automatic Test Cases Generation System [2] it has been used in other
software testing research. Ammann and Offutt argue in [20] that it is a good algorithm for
software testing research because it is easy to understand, relatively small yet very complex in
terms of paths. In 2012 it was used in [21] for research in mutation testing. It has been used in
Master’s Thesis [22] and Doctoral Dissertations [23]. It has been used in ant colony optimization
algorithms [24] and other research [25]. For these reasons the Triangle Classification algorithm
was used in this research.
3. METHODOLOGY
GAs are not always effective at finding optima in multi-optima problems. They can often
converge to local optima, and fail to identify global optima [26-27]. One area of GA research is
called Niche Genetic Algorithms (NGA). NGAs are specialized GAs that find optima in multi-
optima domains [26-27]. These algorithms concentrate on conserving areas of the domain from
the exploitative forces of selection and crossover. This allows the GA to explore these weak
areas of the domain and possibly find local optima.
This research uses the proto-type research methodology and generates a structurally complete set
of test cases for the Triangle Classification algorithm. The results of the program are compared to
existing literature on generating test cases for the Triangle Classification algorithm. The DSGA-
ST algorithm is used with individuals modeling the input parameters for the Triangle
Classification algorithm. Each individual represents a single set of input values.
Our DSGA-ST algorithm produces a list called the tabu list, which contains locally optimal
values. In DSGA-ST the tabu list often contains duplicate or redundant values. So, the tabu list is
filtered through a short algorithm that removes redundant values. These redundant values could
be identical individuals having the same i, j and k values. Or, they could be different individuals,
but that follow the same paths through the program. Figure 3 shows a system level diagram of
how the test cases are generated. When the redundant test cases are removed, the final list of test
cases is the test suite.

128
Figure 2. Control Flow Graph of Triangle Classification Algorithm
Notably, the DSGA [10] is a clustering algorithm framework that uses a tabu list and a radius.
The tabu list stores possible local and global optima and is used to encourage exploration in other
areas of the domain. The radius is used to identify sub-areas of the domain so the algorithm can

129
conserve individuals in these areas. The selection process uses a shared fitness function and tabu
list to encourage exploration in other areas of the domain. DSGA changes its radius as the
algorithm runs. Varying the radius in combination with the tabu list compensates for poor choices
in radius values, which hinder other NGAs.
Figure 3. System Level Diagram of Algorithm
3.1. Algorithm Overview
As with any GA, individuals need to model potential domain values of the problem. To generate a
set of test cases the algorithm models the three input parameters to the Triangle Classification
algorithm, where each represents the length of a side of the triangle. However, the Triangle
Classification algorithm does handle input parameters that cannot form a valid triangle. So, these
other values are possible inputs to the algorithm.
In this research the three sides to the triangle are represented as 8-bit integers with a 9th-bit used
for the sign. Since there are three parameters the total length of an individual is 27 bits. Bytes
were transferred using the little endian representation.
In DSGA [10] a distance measurement needs to be defined between two individuals. In a GA
there are many ways to represent distance. There is Euclidian distance, chromosomal distance and
possibly more ways. In the Triangle Classification algorithm there are a total of 48 edges and
conditions. In this algorithm distance is computed by summing the total number of edges and
conditions that two individuals have in common. The distance is 48 divided by this sum. So two
individuals are considered close together if they share many of the same edges and conditions.
They are considered far apart if they have many different edges and conditions. This distance
measurement will encourage exploration by finding other test cases that explore other parts of the
code.
Table 3 shows the edges and conditions executed in the Triangle Classification algorithm for
examples of (0, 1, 1) and (1, 1, 1). These two individuals have only one edge in common, Edge
1. They have three conditions in common, A1, A2 and A3. That is a total of 4 edges and
conditions. The distance between (0, 1, 1) and (1, 1, 1) is 48 / 4 = 12. However, if we look at (2,
2, 2) it would follow the same path through the graph as (1, 1, 1). So, (1, 1, 1) and (2, 2, 2) would
have all 20 edges and conditions the same and thus would have a distance of 48 / 20 = 2.4. Even
though (0, 1, 1) and (1, 1, 1) share two of the three values, they are far apart in terms of coverage
of the control flow graph. However, (1, 1, 1) and (2, 2, 2) share no values in common but are
identical in the control flow graph.
Table 3. Edges (E) and Conditions for examples (0, 1, 1) and (1, 1, 1)
i = 0
j = 1
k = 1
Edge 1, A1, A2, A3, Edge 2
i = 1
j = 1
k = 1
Edge 1, A1, A2, A3, Edge 3, Edge 4, B1, Edge 5, Edge 6, C1,
Edge 8, Edge 9, D1, Edge 11, Edge 12, E1, Edge 19, G1, Edge
20, Edge 21

130
3.2. Fitness Function
With any GA the fitness function is a function that the GA will find optimal values for. So, the
fitness function should model the problem trying to be solved. In many Multi-optima GAs, like
DSGA, there are two fitness functions. The Natural Fitness is the fitness function that describes
the problem to be solved. The Shared Fitness function [26] will adjust an individual’s Natural
Fitness to encourage exploration in other areas of the domain. DSGA [10] accomplishes this by
decreasing the fitness of individuals the closer that they are to individuals on the tabu list. This is
due to the tabu list already containing possible optimal values. Further exploration around tabu
list values is not helpful. The Shared Fitness encourages exploration in other areas of the domain.
In this research the Natural Fitness is defined as the sum of the number of edges and conditions
covered by the individual. This is done because a goal of structural testing is to find a smallest set
of test cases that cover the structure of the program. The Shared Fitness is defined as the Natural
Fitness divided by mi where mi is defined as the following:
, = - ./ 0 /1 −
0 /1 , 2/345
./ 0 /1 ∗ 0.1
9':
;<) ; => ?
@
(3)
In equation 3 Tabu List Length is the number of individuals on the tabu list. MaxDistance is the
maximum distance between two individuals. In the Triangle Classification algorithm this value is
48. TabuList(x) is the xth individual on the tabu list. Defining mi in this way ensures that as an
individual i gets closer to individuals on the tabu list mi will increase and the Shared Fitness will
decrease.
When DSGA-ST is initialized the tabu list is empty. This makes it difficult to compute Shared
Fitness. So, Shared Fitness is simply the Natural Fitness until individuals are added to the tabu
list.
3.3. DSGA Algorithm – Low Level
DSGA enhances a traditional GA with Seed Selection and Seed Conservation as defined in the
SCGA algorithm [28]. A Seed is defined as a locally strong individual. For each generation of the
GA the individuals are evaluated in order of fitness. An individual is marked as a seed if no other
seed with a radius parameter, r, is a seed. Since the individuals are evaluated in order of fitness
this ensures that the locally strong individuals are seeds. In the example shown in Figure 4 the
seed selection step would evaluate the individuals in the following order: 3, 2, 4, 1, 5, 6. This is
the order of their fitness. It is obvious that individual 3 is a seed. Since individuals 1, 2, 4 and 5
are within the radius of individual 3, they are not designated as seeds. Even though individual 6
is relatively weak, it is outside of the radius of individual 3. When it is evaluated there are no
individuals within the radius that are seeds, so it is designated a seed. In this example individuals
3 and 6 are seeds.

131
Figure 4. Seed Selection Example
Once seeds are determined DSGA goes through the traditional GA steps of Selection, Crossover
and Mutation. The final step in the GA loop is Seed Conservation. Each seed will replace an
individual in the next generation, thus ensuring their genetic traits are passed on. The Seed
Conservation [28] step takes each seed and replaces the weakest individual within the radius, r, of
the seed in the next generation. If no individuals are within the radius, r, of the seed in the next
generation, then the seed replaces the globally weakest seed in the next generation. Once in the
next generation a seed loses its seed designation and needs to go through the Seed Selection
process again to determine if it is a seed. Reference the DSGA pseudo-code for DSGA in Table 4.
Periodically DSGA executes a new step called Reevaluation, which looks for strong areas of the
domain. These areas are placed on the Tabu List. A parameter called Reevaluation Loop Count,
RLC, indicates the number of generations between each Reevaluation step. Individuals can be
placed on the Tabu List in two different ways. The current seeds are placed on the Tabu List. The
Reevaluation step also looks for areas in the population that show convergence. A parameter
called Convergence Limit, CL, is used as a threshold. If in the current generation during
Reevaluation there are CL or more identical individuals then one of the individuals is placed on
the Tabu list. All individuals in the current generation that were placed on the Tabu List are
replaced with randomly generated individuals.
Table 4 shows the DSGA pseudocode. DSGA has shown promise in a number of areas of search
and optimization. In [10] DSGA was used to solve a series of functional optimization problems.
Later in Brown, Pelosi and Dirska [11] DSGA was used to predict stock prices. While selecting
stocks from the Dow Jones Industrial Index the stocks selected outperformed the underlying
index by nearly 400%. Recently it has been used in de novo protein folding problem [12].
Research shows that DSGA is a very good optimization algorithm.
Table 4. DSGA pseudocode [10]
Line # Pseudocode
1 Initialization
2 While not termination condition
3 For (int r = 1; r < RLC; r++)
4 Seed Selection
5 Selection

132
6 Crossover
7 Mutation
8 Seed Conservation
9 End for loop
10 If there exists an individual d with CL or more identical individuals then
11 Add d to tabu_list
12 Replace all individuals identical to d with randomly generated individuals
13 End if
14 Add the seeds of the current generation to the tabu_list
15 Replace all individuals that are seeds with randomly generated individuals
16 Alter radius by SD
17 End while loop
4. RESULTS
We executed the DSGA-ST algorithm with removals of redundancy against the Triangle
Classification algorithm using the parameters shown in Table 5. The individuals, which represent
the three input values for the test cases, were 27-characters long with each character being a 0 or
a 1. There are 1.34 x 108 unique combinations of values that an individual can have. With these
parameters only 200,000 values are actually evaluated. This is less than 0.2% of the total unique
combination of values.
With the parameter settings shown in Table 5, DSGA-ST can produce a suite of tests cases that
cover each statement and all values, TRUE or FALSE, for each condition. The set of test cases is
shown in Table 6 along with the statements covered and values for each of the conditions. This
suite of tests cases executes each statement and produces at least one TRUE and FALSE for each
condition.
Table 5. Parameter values for DSGA-ST
Parameter Description Value
Population Size The number of individuals per generation 200
Number of Generations The number of generations before termination 1,000
Mutation Rate The probability that a gene will mutate 0.03
Initial radius The initial value of the radius 1.0
Radius delta The amount that the radius is changed 0.1
Reevaluation loop count The number of generations needed to evaluate
individuals for the Tabu list
25
Convergence limit The number of identical individuals to conclude that
convergence has taken place
2
Table 6. DSGA Triangle Classification Test Suite
Test # i j k 1 2 3 4 5 6 7 8 9 10 11 12 13 14
1 -117 200 96 X X
2 32 224 64 X X X X X X X
3 40 40 62 X X X X X X X
4 64 218 64 X X X X X X X

133
5 72 -229 -23 X X
6 75 16 16 X X X X X X X
7 90 90 184 X X X X X X X
8 96 32 224 X X X X X X X
9 161 142 140 X X X X X X X
10 196 120 60 X X X X X X X
11 223 174 174 X X X X X X X
12 224 224 224 X X X X X X X X X
13 226 88 -89 X X
14 252 162 252 X X X X X X X
Table 6 (Continuation). DSGA Triangle Classification Test Suite
Test # 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
1
2 X X
3 X X X X
4 X X X X
5
6 X X X X
7 X X X X
8 X X
9 X X
10 X X
11 X X X X X X
12 X X X
13
14 X X X X X
Test # 31 32 A1 A2 A3 B1 C1 D1 E1 F1 F2 F3 G1 H1 H2 I1 I2
1 T F F
2 F F F F F F T F F T
3 F F F T F F F F T T F T
4 X X F F F F T F F F F T T F
5 F T T
6 X X F F F F F T F F F T F T
7 X X F F F T F F F F T F F T
8 F F F F F F T T F F
9 F F F F F F T F F F
10 F F F F F F T F T F
11 F F F F F T F F F T F T
12 F F F T T T F T F T F T
13 F F T
14 F F F F T F F F F T T T

134
Test # J1 J2
1
2
3 F T
4 F T
5
6 T F
7 F T
8
9
10
11 T T
12 F T
13
14 F T
4.1. Discussion of Results
The results show that DSGA is very effective at generating structurally complete test cases. In
only 14 test cases it produced a structurally complete test suite. This is equally as good as the
ATCGS algorithm and better than the GADGET algorithm. A set of 14 test cases is still relatively
small considering the complexity of the algorithm.
Both ATCGS and DSGA-ST produced the same number of test cases at 14. However, in ATCGS
the entire set of test cases is modeled with each individual. For larger programs this could be an
issue. DSGA-ST models each test case as an individual in the GA and the set of test cases comes
from the tabu list with can have an unlimited size. This makes DSGA-ST more flexible.
4. CONCLUSION
This research introduces a new algorithm for automated generation of structural test cases. The
algorithm uses a specialized GA to locate some optimal values and stores them in a short-term
memory structure. As these optimal areas of the domain are discovered, the algorithm encourages
exploration in other areas of the domain. When the GA finishes a step is performed to remove
redundant values, see Figure 3. The remaining test cases make up a complete test suite.
DSGA-ST does not perform any analysis on the target program. It does not make specialized
fitness functions for specific areas of the program. Other than determining which areas of the
target program are covered by a test case, it is completely free of target specific constraints. This
can allow DSGA-ST to generate test cases on much larger programs than the Triangle
Classification algorithm.
Through the use of DSGA-ST and the fitness function, this algorithm generates test cases the way
that many humans would do it by hand. The Natural Fitness function is defined as the sum of the
number of edges and conditions that the test case covers. This encourages the algorithm to find
test cases that cover large areas of the Triangle Classification algorithm. Once some good test
cases are found, the algorithm looks in other areas to find more good test cases.

135
Another benefit that DSGA-ST has over other algorithms is that the number of test cases needed
does not have to be known prior to using the algorithm. In many algorithms the number of test
cases produced cannot exceed the population size of the GA. In DSGA-ST the test cases are
derived from the tabu list which can be larger than the population size.
A future area of this research is to apply DSGA-ST to larger programs and conduct an empirical
study in a software testing organization. This future research should compare defect rates between
software tested through DSGA-ST and other testing methods within the same organization.
When generating structurally complete test cases there are two objectives. The primary objective
is to cover the structure of the program. But a secondary objective is to minimize the number of
test cases. In most GAs all objectives are modeled within the fitness function. But in DSGA-ST
the fitness function only models the number of parts of the program that are covered by a test
case, which is the secondary objective. The primary objective is handled through the tabu list,
distance measurement and the shared fitness. This shows the effectiveness of these components of
the algorithm.
DSGA-ST is a new approach to generating structural test cases. Enhancing a GA with a tabu list,
shared fitness and seeds makes DSGA-ST suitable for finding test cases of large complex
programs. DSGA-ST has a number of advantages over existing algorithms. Future research may
show that it is advantageous in other ways.
REFERENCES
[1] S. Anand, E. Burke, T. Y. Chen, J. Clark, M. B. Cohen, M. B. et. al., "An Orchestrated Survey of
Methodologies for Automated Software Test Case Generation", Journal of Systems and Software
86(8), pp. 1978-2001, 2013.
[2] A. A. Sofokelous & A. S. Andreou, A. S., "Automatic, evolutionary test data generation for dynamic
software testing", The Journal of Systems and Software 81, pp 1883-1898, 2088.
[3] N. K. Gupta & M. K. Rohil, "Using Genetic Algorithm for Unit Testing of Object Oriented
Software", Proceedings of the First International Conference on Emerging Trends in Engineering and
Technology, pp. 308-313, 2008.
[4] J. Wegener & O. Bühler, "Evaluation of different fitness functions for the evolutionary testing of an
autonomous parking system", In Genetic and Evolutionary Computation–GECCO 2004, pp. 1400-
1412, 2004.
[5] M. Pezze & M. Young, Software testing and analysis: process, principles, and techniques. John Wiley
& Sons, 2008.
[6] M. S. Brown, F. Neptune, G. L. Bohl, M. A. Cifuentes, J. Young, et al., "A Survey of the use of
Genetic Algorithms in Structural Testing", International Journal of Advanced Research in Computer
Science and Software Engineering, 3(12), pp. 6-13, 2013.
[7] S. Nachiyappan, A. Vimaladevi, & C. B. SelvaLakshmi, "An Evolutionary Algorithm for Regression
Test Suite Reduction", Proceedings of the International Conference on Communication and
Computational Intelligence, pp. 503-508, 2010.
[8] S. Khor & P. Grogono, "Using a genetic algorithm and formal concept analysis to generate branch
coverage test data automatically", Proceedings of the 19th IEEE International Conference on
Automated Software Engineering, pp. 346-349, 2004.

136
[9] W. Weimer, T. V. Nguyen, C. Le Goues & S. Forrest, "Automatically finding patches using genetic
programming", International Conference on Software Engineering, Vancouver, Canada, pp. 364-374,
2009.
[10] M. S. Brown, (2010). "A Species-Conserving Genetic Algorithm for Multimodal Optimization", PhD.
Dissertation, Nova Southeastern University, Fort Lauderdale, FL, 2010.
[11] M. S. Brown, M. Pelosi & H. Dirska, "Dynamic-radius Species-conserving Genetic Algorithm for
the Financial Forecasting of Dow Jones Index Stocks", Machine Learning and Data Mining in Pattern
Recognition, 7988, pp. 27-41, 2013.
[12] M. S. Brown, T. Bennett & J. A. Coker, "Niche Genetic Algorithms are better than traditional Genetic
Algorithms for de novo Protein Folding", F1000 Research, 3(236) (doi:
10.12688/f1000research.5412.1), 2014.
[13] C. Darwin, On the origin of species by means of natural selection. London: Murray, 1859.
[14] H. J. Bremermann, "The Evolution of Intelligence: The Nervous System as a Model of its
Environment", Technical Report, No.1, Contract No. 477, Issue 17. Seattle WA: Department of
Mathematics, University of Washington, 1958.
[15] K. A. De Jong, W. M. Spears & D. F. Gordon, D. F. "Using Genetic Algorithms for Concept
Learning", Machine Learning 13(2-3), pp. 161–188, 1993.
[16] L. B. Booker, D. E. Goldberg & J. H. Holland, "Classifier Systems and Genetic Algorithms",
Artificial Intelligence 40, pp. 235-282, 1989.
[17] J. H. Holland, Adaptation in Natural and Artificial Systems. Ann Arbor, MI: University of Michigan
Press, 1975.
[18] C. C. Michael, G. McGraw & M. A. Schatz, "Generating software test data by evolution", IEEE
Transactions on Software Engineering 27(12), pp. 1085–1110, 2001.
[19] T. J. McCabe, "A complexity measure", IEEE Transactions on Software Engineering SE-2(4), pp.
308-320, 1976.
[20] P. Ammann & J. Offutt, Introduction to Software Testing. Cambridge University Press, Cambridge,
UK, 2008.
[21] V. H. S. Durelli, J. Offutt & M. E. Delamaro, "Toward Harnessing High-Level language Virtual
Machines for Further Speeding up Weak Mutation Testing", 2012 IEEE Fifth International
Conference on Software Testing. Verification and Validation, pp. 681-690, 2012.
[22] A. Panchapakesan, A Hybrid Genetic Algorithm and Evolutionary Strategy to Automatically
Generate Test Data for Dynamic, White-Box Testing, Master’s thesis, University of Ottawa, 2013.
[23] V. H. S. Durelli, Toward harnessing a Java high-level language virtual machine for supporting
software testing, PhD dissertation, Universidade de São Paulo, 2013.
[24] S. Yang, T. Man, & J. X. "Improved Ant Algorithms for Software Testing Cases Generation", The
Scientific World Journal, 2014.
[25] P. McMinn, "Search based software test data generation: a survey", Software testing, Verification and
reliability, 14(2), pp. 105-156, 2004.
[26] D. E. Goldberg & J. Richardson, "Genetic Algorithms with Sharing for Multimodal Function
Optimization", Proceedings of the Second International Conference on Genetic Algorithms and their
Application, Cambridge Massachusetts, pp. 41-49, 1987.

137
[27] K. A. De Jong, An analysis of the behavior of a class of genetic adaptive systems, PhD. Dissertation,
University of Michigan, 1975.
[28] J. P. Li, M. E. Balazs, G. T. Parks & P. J. Clarkson, "A Species Conserving Genetic Algorithm for
Multimodal Function Optimization", Evolutionary Computation, 10(3), pp. 207-234, 2002.
AUTHORS
Dr. Michael Brown spent nearly 20 years working in the Information Technology
field. Currently he is the Program Chair at University of Maryland University
College for the Software Engineering Master's degree in Information Technology. He
has work for some well-known organizations like Sun Microsystems and NASA.
Dr. Michael Pelois is a faculty member at the University of Maryland University
College and East Central University. He has published numerous papers in the field
of artificial intelligence and the defence industry.

Dynamic Radius Species Conserving Genetic Algorithm for Test Generation for Structural Testing

Recommended

More Related Content

What's hot (17)

Similar to Dynamic Radius Species Conserving Genetic Algorithm for Test Generation for Structural Testing (20)

Recently uploaded (20)

Dynamic Radius Species Conserving Genetic Algorithm for Test Generation for Structural Testing