Wang 2014

ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
The Journal of Systems and Software 000 (2014) 122
Contents lists available at ScienceDirect
The Journal of Systems and Software

journal homepage: www.elsevier.com/locate/jss
Cost-effective test suite minimization in product lines using search

techniques
Shuai Wang a,b,, Shaukat Ali a, Arnaud Gotlieb a
a
b
Certus Software V&V Center, Simula Research Laboratory, P.O. 134, 1325 Lysaker, Oslo, Norway
Department of Informatics, University of Oslo, P.O. 1080, 0316 Blindern, Oslo, Norway
a r t i c l e
i n f o
Article history:
Received 1 December 2013
Revised 9 June 2014
Accepted 16 August 2014
Available online xxx
Keywords:
Product line
Search algorithm
Test suite minimization
a b s t r a c t
Cost-effective testing of a product in a product line requires obtaining a set of relevant test cases from the
entire test suite via test selection and minimization techniques. In this paper, we particularly focus on test
minimization for product lines, which identies and eliminates redundant test cases from test suites in order
to reduce the total number of test cases to execute, thereby improving the eciency of testing. However, such
minimization may result in the minimized test suite with low test coverage, low fault revealing capability,
low priority test cases, and require more time than the allowed testing budget (e.g., time) as compared to the
original test suite. To deal with the above issues, we formulated the minimization problem as a search problem
and dened a tness function considering various optimization objectives based on the above issues. To assess
the performance of our tness function, we conducted an extensive empirical evaluation by investigating
the tness function with three weight-based Genetic Algorithms (GAs) and seven multi-objective search
algorithms using an industrial case study and 500 articial problems inspired from the industrial case study.
The results show that Random-Weighted Genetic Algorithm (RWGA) signicantly outperforms the other
algorithms since RWGA can balance all the objectives together by dynamically updating weights during
each generation. Based on the results of our empirical evaluation, we also implemented a tool called TEst
Minimization using Search Algorithms (TEMSA) to support test minimization using various search algorithms
in the context of product lines.
2014 Elsevier Inc. All rights reserved.
1. Introduction
Product Line Engineering (PLE) has proven to be a cost-effective
way to develop similar products by exploiting and managing commonalities and variabilities among a large number of products
(Benavides et al., 2010; McGregor, 2001). PLE has shown promising
benets in both academia and industry such as reducing development
time and costs, speeding up product time-to-market and improving
the quality of products of a product line family (Benavides et al.,
2010). Cost-effective testing of a new product in a product line can be
achieved by systematic application of model-based PLE approaches
(McGregor, 2001), for example, based on feature models (Benavides
et al., 2010). Feature models have become the de-facto standard for
capturing commonalities and variabilities in product lines and can
be used for systematic selection of test cases for a new product in a
product line (Benavides et al., 2010).
Corresponding author at: Certus Software V&V Center, Simula Research Laboratory,
P.O. 134, 1325 Lysaker, Oslo, Norway. Tel.: +47 45239921.
E-mail addresses: shuai@simula.no (S. Wang), shaukat@simula.no (S. Ali),
arnaud@simula.no (A. Gotlieb).
An overall view of our product line testing project is presented

in Fig. 1. The project has two main activities: (1) test selection for
a product in a product line and (2) test minimization for the selected test cases following the rst activity. Test selection was focused in our previous works (Wang et al., 2012, 2013a), where we
proposed a methodology for automated selection of test cases for a
new product using Feature Model (FM) and Component Family Model
(CFM) (GmbH, 2006). FM was used to capture commonalities and
variabilities of features of a product line, where CFM was used to
model the structure of test cases in a test case repository as shown in
Fig. 1. Two types of traceability links were established: (1) traceability links (traces) from CFM to actual test cases in the repository; (2)
links between FM and CFM using pre-dened restrictions. To test a
new product (e.g., Producty as shown in Fig. 1), a test engineer only
requires performing simple selection of relevant features in the FM
and the corresponding test cases can be automatically obtained from
the test case repository.
Based on our experience of working with Cisco, we discovered that even after test selection, the number of test cases is still
large (more that 2000 requiring 34 days to execute per product) and many of the test cases are redundant (e.g., several test
cases cover the same functionality). On top of that, new test cases
http://dx.doi.org/10.1016/j.jss.2014.08.024
0164-1212/ 2014 Elsevier Inc. All rights reserved.
Please cite this article as: S. Wang et al., Cost-effective test suite minimization in product lines using search techniques, The Journal of Systems
and Software (2014), http://dx.doi.org/10.1016/j.jss.2014.08.024
JID: JSS
2
ARTICLE IN PRESS
[m5G;December 8, 2014;18:52]
S. Wang et al. / The Journal of Systems and Software 000 (2014) 122
Fig. 1. An overview of cost-effective testing of products in a product line.
for product lines are continuously added further increasing the number of test cases. Executing all these test cases is practically impossible within the allocated testing time and thus warrants efcient and sophisticated test minimization techniques (Engstrm
and Runeson, 2011; Harman, 2011; Lopez-Herrejon et al., 2013;
Runeson and Engstrm, 2012; Wang et al., 2013c). However, such
minimization may reduce cost (e.g., execution time), but on the
other hand may also reduce effectiveness (e.g., fault detection capability). This means an ecient technique must achieve a desirable
balance among cost and effectiveness measures, which is a multiobjective optimization problem. A number of such objectives have
been studied for multi-objective test optimization in regression testing in the existing literature such as execution time, test coverage and fault detection capability (Harman, 2011; Yoo and Harman,
2012).
Multi-objective search algorithms are well adapted for the type
of problem we are solving in this paper (Harman et al., 2009; Konak
et al., 2006). Moreover, Weight-based Genetic Algorithms (GAs) are
also known to solve multi-objective optimization problems by assigning a set of weights to each objective (Konak et al., 2006). In
our previous work (Wang et al., 2013c), we dened a tness function based on the following cost/effectiveness measures based on
our industrial collaboration and the existing literature (Engstrm and
Runeson, 2011; Harman, 2011; Lopez-Herrejon et al., 2013): (1) Cost:
minimizing the number of test cases thereby indirectly reducing execution cost referred as Test Minimization Percentage (TMP); (2) Effectiveness: High Feature Pairwise Coverage (FPC) and Fault Detection
Capability (FDC). In Wang et al. (2013c), we compared three weightbased Genetic Algorithms (GAs) and showed that Random-Weighted
GA (RWGA) had signicantly better performance than the other
algorithms.
Further investigation at Cisco revealed the following additional
cost and effectiveness measures: (1) Cost: overall execution cost of
the minimized test cases; (2) Effectiveness: priority of test cases.
Notice these two additional measures have also been investigated
in the existing literatures (Engstrm and Runeson, 2011; Harman,
2011; Lopez-Herrejon et al., 2013). Moreover, we also decided to
compare weight-based algorithms with other types of algorithms
such as Fast Non-dominated Sorting Genetic Algorithm (NSGA-II)
(Konak et al., 2006; Yoo and Harman, 2007). Based on these observations, this paper extends our existing work (Wang et al., 2013c)
by including: (1) four effectiveness measures, i.e., Test Minimization
Percentage (TMP), Pairwise Coverage (FPC), Fault Detection Capability
(FDC) and Average Execution Frequency (AEF) and one cost measure,
i.e., Overall Execution Time (OET); (2) multi-objective tness function considering all the proposed cost/effectiveness measures; and
(3) empirical evaluation for the proposed tness function in conjunction with three weight-based multi-objective search algorithms and
seven other multi-objective search algorithms on an industrial case
study and assessing the scalability using 500 articial problems of
varying complexity; (4) a tool called TEst Minimization with Search
Algorithms (TEMSA) based on Durillo and Nebro (2011).
The results of our empirical evaluation showed that: (1) based
on the industrial case study, all the selected search algorithms except Improved Strength Pareto Evolutionary Algorithm (SPEA2) are
cost-effective for solving the test minimization problem and RWGA
achieves the best performance; (2) for the 500 articial problems, the
results are consistent with the industrial case study, i.e., RWGA outperforms all the other search algorithms. Furthermore, RWGA managed to solve the problems of varying complexity.
The rest of the paper is organized as follows. Section 2 provides a brief introduction of the selected search algorithms. Section 3
presents a formal representation of our problem, denitions and functions for the cost/effectiveness measures and tness function used by
all the algorithms. Section 4 describes our industrial case study and
designed articial problems followed by empirical evaluation of the
applied algorithms in Section 5. Section 6 discusses the results and
analysis for the experiments and Section 7 presents an overall discussion based on the obtained results. In Section 8, the tool support
is discussed. Section 9 addresses the threats to validity of our empirical study and related work is discussed in Section 10. Section 11
concludes the paper and presents future work.
2. Description of the selected algorithms
In this section, we provide a brief description of the ten algorithms that were selected for our experiments. These algorithms
are classied into four working mechanisms to guide the search
ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
Table 1
Classication of the selected search algorithms.
Different mechanisms
Weight-Based GA
GAs
Evolutionary
Algorithms (EAs)
Swarm Algorithm
Hybrid Algorithm
Stochastic Algorithm
Sorting-Based GA
Cellular-Based GA
Strength Pareto EA
Evolution Strategies
Particle Swarm Optimization
Cellular genetic algorithm + differential evolution
Random Search
(Brownlee, 2012), which are: Evolutionary Algorithms (EAs), Swarm

Algorithms, Hybrid Algorithms and Stochastic Algorithms (Table 1).
The column Minimization Before shows whether the listed algorithms have been adapted for solving the test minimization problem in the existing literature. Notice that one of our aims is to
evaluate a representative set of the existing search algorithms rather
than limiting us to the commonly used algorithms (e.g., NSGA-II and
SPEA2). To achieve this goal, we chose ten search algorithms systematically based on the categories classied in (Brownlee, 2012). Before
getting into more details for each algorithm (Sections 2.22.5), we
rst provide a brief overview for search algorithms (Section 2.1).
2.1. Overview of search algorithms

Generally speaking, search algorithms aim at mimicking natural
phenomenon, for example, natural evolution process or behaviors of
animals (e.g., bird ocking) to search optimal solutions for various
optimization problems (Brownlee, 2012). To apply search algorithms,
tness function (objective function) should be dened mathematically to guide the search with aims at assessing the capability of the
solutions.
Moreover, for genetic algorithms (GAs), genes, individuals (also
called chromosomes) should be carefully dened beforehand. More
specically, individuals refer to potential solutions with respect to
the optimization problem and each individual (solution) is composed of a set of genes, which present units for the solution, e.g.,
test cases for the test suite in our case. The set of individuals handled
by GAs is represented as population. Similarly, for swarm algorithm
(Section 2.3), the potential solutions are named as particles and the
set of the solutions handled by algorithms refers to swarm.
After properly dening tness function (objective function), the
search algorithms will be repeatedly run by applying various operators (e.g., selection) to produce the best individual (solution) until
the terminating condition is satised (e.g., number of tness evaluation). Three operators are usually adapted by various search algorithms, namely selection, crossover (also called recombination) and
mutation (Brownlee, 2012). More specically, (1) Selection operator
corresponds to the operation of selecting the solutions (e.g., individuals for evolutionary algorithms) with the best values based on the
dened tness function. Notice that the Selection operator can be applied to all the search algorithms; (2) Crossover operator is mainly
applied in various GAs and refers to the operation of selecting two
individuals (usually called parent individuals) and exchanging their
genes partially. The genes that will be switched between the individuals are usually determined at random. New individuals obtained
by crossover operator are usually called offspring; and (3) Mutation
operator is also mostly adapted in GAs and refers to updating the individuals by changing the properties of genes randomly based on the
predened mutation rate (i.e., probability of changing the properties
of genes).
Algorithms
Minimization before?
WBGA
WBGA-MO
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
Yes
Yes
Yes
Yes
No
Yes
No
No
No
Yes
2.2. Evolutionary Algorithms (EAs)

EAs are inspired by biological evolution, which was rst proposed
by Darwin to describe how natural selection and mutation occurs on
various genes of species (Brownlee, 2012). Since EAs are well known in
search-based software engineering (Harman et al., 2009), we choose
seven of them, which are further classied into GAs, Strength Pareto
EA and Evolution Strategies. We describe each of them in detail as
below.
The working theory of these three weight-based GAs is to assign a
particular weight to each objective function for converting the multiobjective problem to a single objective problem with a scalar objective function (Konak et al., 2006). The difference of these algorithms
can be summarized as the different mechanisms of assigning weights
for each objective. Weight-Based GA (WBGA) assigns xed weights
for each objective dened for each objective beforehand, which can
be provided by users based on the domain knowledge and expertise. The pre-determined weight for each objective is used during all
generations of the algorithm. WBGA for Multi-objective Optimization (WBGA-MO) uses a weight pool including a set of predened
weights and each solution can choose weights for each objective randomly at each generation. The assigned weight for each objective
changes during each generation, but is still limited to the ones in the
weight pool. Notice that such weight pool can also be pre-provided by
users. RWGA randomly assigns normalized weights to multiple objective functions for each solution when selecting ttest individuals
at each generation, i.e., assigning weights to objective functions is dynamic and there is no need to set xed weights for different objectives
beforehand.
NSGA-II is based on the Pareto dominance theory, which outputs
a set of non-dominated solutions for multiple objectives (Deb et al.,
2002). The population is sorted into several non-dominated fronts using a ranking algorithm rst. Then, individual solutions are selected
from these non-dominated fronts by calculating the crowd distance,
which is used to measure the distance between the individual solutions and the rest of the solutions in the population (Konak et al.,
2006). If two individual solutions are in the same non-dominated
front, the solution with a higher value of crowd distance will be
selected.
Multi-objective Cellular (MOCell) is based on the cellular model of
GAs (cGAs) with an assumption that an individual only interact with
its neighbors during the search process (neighborhood) (Nebro et al.,
2007). Moreover, MOCell stores a set of obtained non-dominated individual solutions in an external archive. After each generation, MOCell
replaces a xed number of solutions chosen randomly from the population by selecting the same number of solutions from the archive
until the termination conditions are met. Such replacement only takes
place when the newly generated solutions from the population are
worse than the solutions in the archive.
Improved Strength Pareto Evolutionary Algorithm (SPEA2) is also
based on Pareto dominance theory (Zitzler et al., 2001). The tness
JID: JSS
4
ARTICLE IN PRESS
[m5G;December 8, 2014;18:52]
value for each solution is calculated by summing up a strength raw

tness based on the dened objective functions and density estimation, where density estimation measures the distance between a
solution and its nearest neighbors for maximizing the diversity of the
obtained solutions. SPEA2 rst stores a xed number of best solutions
into an archive by applying selection, crossover and mutation operators. Then, a new population is created by combining the solutions
from the archive and the non-dominated solutions of the original solution. If the combined non-dominated solutions are more than the
maximum size of population, the solution with minimum distance to
any other solutions is selected by applying a truncation operator to
calculate the distances to the its neighborhood.
Pareto Archived Evolution Strategy (PAES) is inspired by evolution
theory (Knowles and Corne, 2000). By applying dynamic crossover,
mutation operators, PAES aims at maximizing the suitability for
the selected solutions. Moreover, PAES stores the obtained nondominated solutions into an archive and newly generated solutions
can be added into such archive if they are better than the existing
solutions by calculating objective functions.
formal representation of the search problem, followed by denitions

and functions for the cost/effectiveness measures in Section 3.2. In
Section 3.3, the overall tness function is dened.
2.3. Swarm Algorithm
product pi which is to be tested (Benavides et al., 2010), where fi

can be any feature in F (fi F ) and nfpi is the number of features of
Fpi (1 nfpi nf ).
TSpi = {t1 , t2 , t3 , . . . , tntp } is a subset of the test suite TS for testing
i
the pi product, where ti can be any test case in TS (ti TS) and 1
ntpi nt. Notice in our case, TSpi is obtained by our pervious test case
selection methodology proposed in Wang et al. (2012, 2013a). Notice
that the proposed approach in this paper can be applied separately
(without the test case selection methodology) for multi-objective test
minimization problems in other contexts.
Spi = {Spi 1 , Spi 2 , Spi 3 , . . . , Snsp } is a set of potential solutions based
i
on TSpi for pi , where nspi is the total number of solutions for testing pi , which can be measured as 2ntpi 1. Notice that each potential solution includes a set of test cases from TSpi . As the number of
test cases increases, the potential solutions increase exponentially.
Suppose we have 1000 test cases to test the product line P and using the methodology proposed by Wang et al. (2012, 2013a), 100
test cases are obtained to test the product pi . Then there will be
2100 1 potential solutions for testing pi , whose solution space is
huge.
3.1. Problem representation

To precisely dene our test minimization problem, the following
denitions are presented rst.
3.1.1. Basic concepts

Let P = {p1 , p2 , p3 , . . . , pnp } be a product line with a set of products,
where np is the number of products in P.
F = {f1 , f2 , f3 , . . . , fnf } is a feature model with a set of features to
represent P (Benavides et al., 2010), where nf is the number of features,
i.e., functionalities need to be tested for P.
TS = {t1 , t2 , t3 , . . . , tnt } is a test suite including a large number of
test cases (nt) for testing P.
Fpi = {f1 , f2 , f3 , . . . , fnfp } is a subset of F to represent a specic
i
Particle Swarm Optimization (PSO) is a biological metaheuristic

inspired by the social foraging behavior of animals such as bird ocking (Brownlee, 2012).
Speed-constrained Multi-objective Particle Swarm Optimization
(SMPSO) algorithm is based on PSO theory (Knowles and Corne, 2000;
Nebro et al., 2009). Similar as NSGA-II, SMPSO selects best solutions
by calculating crowding distance and also stores the selected individual solutions in an archive. SMPSO takes advantage of mutation
operators to accelerate the speed of convergence and adapts velocity
constriction mechanism to avoid the explosion of swarms. According
to Knowles and Corne (2000) and Nebro et al. (2009), the performance of SMPSO is considered as the best one among the existing
PSO algorithms.
2.4. Hybrid Algorithm
Differential Evolution (DE) algorithm is considered as another kind
of EA, which generates solutions by applying recombination, mutation, and selection operators (Brownlee, 2012). More specically, DE
calculates the weighted difference between two randomly selected
solutions and integrate the obtained parts into a third solution for
generating a new solution.
CellDE is a hybrid metaheuristic algorithm using MOCell as a
search engine and replacing the typical selection, crossover and mutation operators for GAs with the recombination mechanism of DE
(Durillo et al., 2008). It is considered that CellDE takes the advantage of cellular GA and DE with good diversity and convergence and
the performance of CellDE improves signicantly as compared with
separate ones (Durillo et al., 2008).
2.5. Stochastic Algorithm
Random Search (RS) is a stochastic algorithm and is commonly
used as a baseline for evaluation (Brownlee, 2012). RS is adapted to
randomly generate solutions from the entire search space during each
generation. Notice each newly generated solution is independent of
the obtained solutions.
3. Problem representation and tness function
To guide the search toward an optimal solution using multiobjective search algorithms, it is essential to dene a tness function,
as is the case for any search algorithm. In Section 3.1, we present a
3.1.2. Cost/effectiveness measures

Cost = {Cost1 , Cost2 , . . . , Costncost } is a set of cost measures for
minimization, e.g., execution time, cost of a test minimization algorithm.
Cost (si , CostMeasure) is a function, which returns the cost of a
solution si from Spi based on a cost measure (CostMeasure) from Cost.
Eff ect = {eff ect1 , eff ect2 , . . . , eff ectneff ect } is a set of effectiveness
measures for minimization, e.g., test minimization percentage, feature pairwise coverage.
Eff ect (si , Eff ectivenessMeasure) is a function, which returns the effectiveness of a solution si from Spi based on an effectiveness measure
(EffectivenessMeasure) from Effect.
Each solution si in Spi comprised ntsi test cases from TSpi :
{tsi 1 , tsi 2 , tsi 3 , . . . , tntsi }, where 1 ntsi ntpi and has a set of values for cost measures Cost (si , CostMeasure) and effectiveness measures Eff ect (si , Eff ectivenessMeasure), which are dened in the next
section.
3.1.3. Problem representation

Our optimization problem can be represented as:
Problem. Search
for
solution
sk
(sk
comprised
{tsk 1 , tsk 2 , tsk 3 , . . . , tntsk } from TSpi , where 1 ntsk ntpi ) from
ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
Spi for testing the product pi to achieve the following two objectives
(i.e., maximum effectiveness and minimum cost):
sl Spi sl = sk :
neff ect
j=1
ncost
neff ect
Eff ect (sk , eff ectj )
Eff ect (sl , eff ectj )
j=1
Cost (sk , cos tj )
j=1
ncost
Cost (sl , cos tj )
3.2. Denitions and functions for effectiveness/cost measures

In this section, based on the above-mentioned problem, we provide mathematical denitions and functions of relevant effectiveness
and cost measures in our context.
3.2.1. Effectiveness measures
In our context, effectiveness measures Effect include four elements, i.e., test minimization percentage (TMP), feature pairwise coverage (FPC), fault detection capability (FDC) and average execution
frequency (AEF). We dene them in detail below:
3.2.1.1. Test minimization percentage (TMP). Denition 1. TMP is to
measure the amount of reduction for the number of test cases.
Objective Function 1. TMP is calculated as follows:
ntsk
TMPsk = 1
ntpi
100%
As discussed in Section 3.1, ntsk is the number of test cases for

the specic solution sk obtained by search algorithms for testing the
product pi , where 1 ntsk ntpi . ntpi is the number of test cases in
the test suite for testing product pi . TMP value ranges from 0 to 1
and a higher value of TMP shows more test cases are reduced in the
minimized test suite.
3.2.1.2. Feature pairwise coverage (FPC). Denition 2. FPC is to measure how much pairwise coverage can be achieved by a chosen solution (Kuo-Chung and Yu, 2002).
We chose this type of coverage based on our domain knowledge,
discussion with test engineers, and history data about faults since a
higher percentage of detected faults are mainly due to the interactions between features. Notice that other types of coverage (e.g., code
coverage) can also be introduced into the tness function based on
other cases. FPC in our case is designed to compute the capability of
covering feature pairs by a chosen solution.
Objective Function 2. The function for measuring FPC is shown as
below.
FPCsk =
NumFPsk
NumFPpi
Num_FPsk is the number of feature pairs covered in the test cases

for the solution sk measured as follows.
nt
Num_FPsk =
sk
test three features. Then, the feature pairs covered by test case tci
are C23 = 3 2/2 = 3. Notice that if some feature pairs are repeated
ones compared with the feature pairs covered by the previous test
cases, repeated pairs will be removed when computing NumFPs .
k
Meanwhile, feature pairs only refer to the valid feature pairs that
are not violated against the constraints dened among various features. In our case, there are no specic constraints that are dened
on the features. But if such constraints exist in our contexts, the invalid feature pairs will be eliminated when calculating NumFPs and
k
NumFPp .
i
Num_FPpi is all number of feature pairs for testing the product pi
size(Fp )
j=1
Num_FPtci
i
measured as: Num_FPpi = C2
= ntpi (ntpi 1)/2. Fpi is the set
of features representing the product pi including ntpi features. For
instance, if pi is represented by ten features, all feature pairs covered
by pi are C210 = 45. Notice FPC is calculated for a chosen test solution
ranging from 0 to 1 and a higher value shows better feature pairwise
coverage.
3.2.1.3. Fault detection capability (FDC). Denition 3. FDC is to measure the fault detection capability of a selected test solution for a
product.
In our context, fault detection refers to the rate of successful execution for a test case in a given time, e.g., a week in our case. More
specically, the execution of a test case can be dened as a success if
it can detect faults in a given time (a week in our case) and as a fail if
it does not detect any fault.
Objective Function 3. FDC is measured using the following function.
ntsk
SucRtci
ntsk
i=1
FDCsk =
sk refers to a solution for a product including a set of test cases and

ntsk is the number of test cases for the solution sk , where 1 ntsk
ntpi . SucRtci is the successful rate of execution for test case tci , which
can be measured as below.
SucRtci =
NumSuctci
NumSuctci + NumFailtci
NumSuctci is the number of success for test case tci and NumFailtci
is the number of fail for test case tci . For instance, a test case is usually
executed 1000 times per week in Cisco. So if it executes successfully
for 800 times, the SucR is 800/1000 = 0.8.
Note that FDC value also ranges from 0 to 1 and a higher value
represents better fault detection capability.
3.2.1.4. Average execution frequency (AEF). Denition 4. AEF is to measure the average execution frequency of the solution sk based on the
execution frequency of each included test cases during a given time,
e.g., a week.
In our context, we found that a test suite may have high priority if it is executed very frequently during the given time based on
the domain knowledge. Therefore, the minimized test suite should
preferably consist of test cases with high priority, which can be determined by the average execution frequency. This is also our main
objective to seek a solution with higher average execution frequency
from the solution space.
Objective Function 4. The function to measure AEF is presented as
below.
ntsk
i=1
ntsk is the number of test cases for the solution sk , where 1

ntsk ntpi . Num_FPtci is the number of unduplicated feature pairs
covered by the test case i(tci ). The feature pairs covered by tci can be
size(Ftc )
i
computed as: Num_FPtci = C2
size(Ftci ) is the number of features tested by test case tci . For instance, test case tci is used to
AEFsk =
EFtci
ntsk
i=1
ntsk is the number of test cases for the solution sk , where 1

ntsk ntpi . EFtci is the execution frequency for the test case tci in a
given week. For instance, suppose the solution sk consists of three
ARTICLE IN PRESS
JID: JSS
6
test cases: tc1 , tc2 and tc3 . These three test cases are executed 5
times, 6 times and 7 times, respectively. Then AEF for the solution
sk will be 6 (i.e., (5 + 6 + 7)/3 = 6). Notice that a higher value of
AEF represents higher average execution frequency thereby higher
priority.
3.2.2. Cost measures
Cost measures Cost includes one element in our context, i.e., overall
execution time (OET), which is dened as below.
3.2.2.1. Overall execution time (OET). Denition 5. OET is to measure
the overall execution time for the solution sk consisting of a set of test
cases.
Objective Function 5. OET can be measured using the function as
below.
nt
OETsk =
[m5G;December 8, 2014;18:52]
sk
AETtci
i=1
ntsk is the number of test cases for the solution sk , where 1 ntsk
ntpi . AETtci is the average execution time for test case tci in the solution
sk , which can be calculated as below.
EFtci
AETtci =
ETtci
EFtci
i=1
EFtci is the execution frequency of the test case tci in a given

week and ETtci is the overall time for each execution. For instance,
the frequency of the test case tci is three in a given week and the
time for each execution is 15 min, 20 min and 25 min. Then the
AETtci = (15 + 20 + 25)/3 = 20 min. Notice that a higher value of
OETsk shows more time is required for the solution sk to execute
and our goal is to keep such time within the allowed time budget.
In summary, our test minimization problem is formulated as a veobjective optimization problem, which can be solved using search
algorithms.
3.3. Fitness function
Based on the dened effectiveness measures and cost measures in
the previous section, our tness function is presented as below. Notice
that a lower value of tness function represents better performance
for a solution.
Fitness = f (TMP, FPC, FDC, OET, AEF )

Notice that the obtained values by different objective functions
do not stay at the same comparable level. For instance, FDC ranges
from 0 to 1 and the value of OET can be 1.5 (h). So we rst normalize the obtained values of all the ve functions using the following
normalization function (Greer and Ruhe, 2004; Zhang et al., 2008).
nor(x) =
x
x+1
where x can be any value obtained for each objective (i.e., TMP, FPC,
FDC, OET and AEF), such as 0.56 for FDC and 0.5 h for OET. Therefore,
the specic tness function for the weight-based GAs (i.e., WBGA,
WBGA-MO and RWGA) can be formulated as below, which ranges
from 0 to 1, where lower the value is, better the tness function is.
Fitness_w = 1 w1 nor(TMP ) w2 nor(FPC ) w3

nor(FDC ) w4 (1 nor(OET )) w5 nor(AEF )
w1 , w2 , w3 , w4 and w5 are the weights assigned to TMP, FPC,
FDC, OET and AEF, respectively satisfying the following constraint:
5
i=1 wi = 1. Using this way, multi-objective optimization problem is
Table 2
Four products in Saturn.
Product
# features
# test cases
C20
C40
C60
C90
17
25
32
43
138
167
192
239
converted to a single objective problem with a scalar objective function, which is a classical approach and is ecient to be solved using
GAs (Konak et al., 2006). Notice that RWGA assigns weights to objectives dynamically during each generation and it is not required to set
xed weights to different objectives before.
The other search algorithms are based on different mechanisms
and there is no need to assign particular weights to each objective
function. All the cost/effectiveness measures including ve objectives
with functions have been clearly dened in Section 3.2. Notice that
using NSGA-II, MOCell, SPEA2, PAES, SMPSO, CellDE, a set of individual solutions can be obtained after a specic number of tness
generations.
4. Industrial case study and articial problems
First, we used an industrial case study of Videoconferencing System (VCSs) product line called Saturn provided by Cisco Systems,
Norway to evaluate our tness function (CiscoSystems, 2010). Saturn has several VCS products including C20 and C90 and has more
than 2000 test cases. Notice that not all test cases are relevant for
each product and new test cases are developed for Saturn every
day. Based on our collaboration with Cisco, we observed that the
current testing practice is performed in two ways: (1) executing
all the test cases for a product; (2) randomly selecting a subset of
test cases for a product. Therefore, random search could be considered as an equivalent case, though not a very accurate measure
and require additional experiments involving real testers, which can
be performed in the future. The following paper will evaluate the
performance of the selected search algorithms as compared with
RS.
We chose four products C20, C40, C60 and C90 from Saturn. There
are 169 features in Saturn and each product includes a subset of these
features. Meanwhile, each feature can be tested by at least one test
case (usually more than one). More details are shown in Table 2. For
instance, C20 has 17 features and 138 test cases are used to test these
features. Each test case tci has a successful rate for execution (SucRtci ),
an average execution time (AETtci ) and an execution frequency (EFtci )
(Section 3.2). In general, for Saturn, each feature is associated with
510 test cases and each test case tci is associated with 15 features with SucRtci ranging from 50% to 95%, AETtci ranging from 2
min to 60 min and EFtci ranging from 1 time per week to 50 times
per week.
Moreover, we dened 500 articial problems to empirically evaluate whether the tness function dened in Section 3 can address
our test minimization problem even with a large number of features
and test cases. With this objective in mind, we rst created a feature repository including 1200 features and a test case repository
of 60,000 test cases. For each test case tci , three key attributes are
assigned randomly (inspired by our industrial case study but with
expansion for generality), i.e., SucRtci ranges from 0% to 100%, AETtci
ranges from 1 min to 100 min and EFtci ranges from 1 time to 100
times per week. Moreover, we created articial problems with the
increasing number of features and test cases, i.e., we used a range
of 101000 with an increment of 10 for features number and each
feature can be associated with test cases ranging from 5 to 45 with an
increment of 10. So that 100 5 = 500 articial problems were obtained in this way. Based on that, each articial problem consists of a
JID: JSS
ARTICLE IN PRESS
[m5G;December 8, 2014;18:52]
set of features associated with a test suite including a set of test cases
and the goal is to minimize the test suite using the selected multiobjective algorithms.
5. Empirical evaluation
This section rst presents experiment design including: (1) research questions required to be addressed, (2) selected criteria for
the algorithms and parameter settings, and (3) evaluation mechanisms for all the search algorithms. Moreover, we also describe relevant statistical tests used for our experiments followed by experiment
execution.
5.1. Experiments design
The goal of our experiments is to evaluate the tness function in
conjunction with the ten search algorithms (i.e., WGBA, WBGA-MO,
RWGA, NSGA-II, MOCell, SPEA2, PAES, SMPSO, CellDE and RS) in terms
of addressing the optimization problem: minimizing the test suite for
testing a product with high TMP, FPC, FDC and AEF, and OET within
allocated time budget.
5.1.1. Research questions
We will answer the following research questions with our experiments:
1. RQ1: Are the selected multi-objective search algorithms
cost/effective to solve our test minimization problem?
2. RQ2: Are the selected multi-objective search algorithms
cost/effective as compared to RS?
3. RQ3: Which one achieves the best performance among the selected multi-objective search algorithms?
4. RQ4: How does the increment of the number of features and
test cases impact the performance of the selected multi-objective
search algorithms?
5.1.2. Selected criteria for the algorithms and parameter settings
In our experiments, we rst compared three weight-based multiobjective GAs and RS, i.e., WGBA with two set of xed weights
based on the domain knowledge and expertise WBGA_W1 (W1 =
(0.2, 0.2, 0.2, 0.2, 0.2)), WBGA_W2 (W2 = (0.1, 0.2, 0.2, 0.4, 0.1)),
WBGA-MO and RWGA. For all of them, we used a standard onepoint crossover with a rate of 0.9 and mutation of a variable is done
with the standard probability 1/n, where n is the number of variables (defaulted standard parameter settings from jMetal (Durillo and
Nebro, 2011)). Meanwhile, the size of population and maximum number of tness evaluation are set as 100 and 25,000, respectively (all
these standard settings performed well in our previous work (Wang
et al., 2013c)). Finally, RS was used as the comparison baseline to
assess the diculty of the addressed minimization problems (Ali
et al., 2010).
Moreover, we compared the best weight-based GA with the other
six selected search algorithms, i.e., NSGA-II, MOCell, SPEA2, PAES,
SMPSO, CellDE (RS is still used as a baseline for assessing the difculty of the problems (Ali et al., 2010)). For all the other algorithms, we also used the default suggested parameter settings from
jMetal (Durillo and Nebro, 2011) as shown in Table 3. Notice that
different settings may lead to different performance for genetic algorithms, but standard settings usually perform well (Arcuri and Briand,
2011).
5.1.3. Evaluation mechanisms
To address RQ1, a set of thresholds for TMP, FPC, FDC and OET
are selected showing the minimum acceptable values for a particular context. Notice that these thresholds are set through the domain analysis, discussion with test engineers and a test manager
Table 3
Parameterization for the selected search algorithms (n is the number of variables).
Parameterization used in NSGA-II and SPEA2
Population Size
100
Selection of Parents
Binary tournament + binary tournament
Recombination
Simulated binary, crossover rate = 0.9
Mutation
Polynomial, mutation rate = 1.0/n
Parameterization used in MOCell
Population Size
100
Neighborhood
1-hop neighbors (8 surrounding solutions)
Recombination
Simulated binary, crossover rate = 0.9
Mutation
Archive Size
100
Parameterization used in PAES
Mutation
Archive Size
100
Parameterization used in SMPSO
Swarm Size
100
Mutation
Archive Size
100
Parameterization used in CellDE
Population Size
100
Neighborhood
1-hop neighbors (8 surrounding solutions)
Recombination
Differential evolution, crossover rate = 0.9
Archive Size
100
at Cisco, and test execution history. Meanwhile, there is no specic

threshold provided for AEF since the test engineers think that it is
not meaningful for them to propose such threshold. In our context,
these thresholds values are: TMP := 0.8; FPC := 0.8; FDC := 0.85 and
OET := 2 (h). Notice the purpose of these pre-dened thresholds is
to assess the performance of the selected search algorithms and the
obtained solutions are independent of the availability of these threshold values. In other contexts, experts can provide such thresholds if
required.
As discussed in Section 2, NSGA-II, MOCell, SPEA2, PAES, SMPSO,
CellDE may produce more than one solutions. Our evaluation mechanism is to choose the best solution with the highest value of objective
functions for TMP, FPC, FDC and OET and compare the values of objective functions with provided thresholds separately. For instance,
for TMP using NSGA-II, the highest value of objective function for the
best solution is rst chosen from the obtained ones. The value of the
objective function is then compared with the threshold for TMP. Notice that for each objective, the chosen solution may be different so
that it is essential to seek a mechanism to evaluate the overall performance of obtained solutions by the algorithms considering all the
objectives together. Our mechanism is to calculate the average value
to combine the ve provided thresholds and the values of objective
functions for the solutions obtained by all the search algorithms. The
formula for combining is shown as below and the combined value is
called as Overall Fitness Value (OFV).
OFV =
(nor(TMP) + nor(FPC ) + nor(FDC ) + (1 nor(OET )) + nor(AEF ))

5
Afterwards, we can evaluate the overall performance for all the

selected search algorithms using the values of OFV. Such mechanism
makes sense since we dene the same objective functions, and use
the same formula to combine the values of objective functions for
all the algorithms. Notice that higher the OFV value is, better the
performance of a specic algorithm is.
To answer RQ2, RQ3 and RQ4, the same mechanisms are adapted
as above when comparing all the algorithms. Notice that for the
industrial case study, we rst compare and analyze the results obtained by the selected algorithms for each objective separately and
then evaluate the OFV considering the ve objectives together. As
for articial problems, we only evaluate the values of OFV to assess the performance and scalability for the ten algorithms. Moreover, for articial problems, Mean Fitness Value for each problem
ARTICLE IN PRESS
JID: JSS
8
[m5G;December 8, 2014;18:52]
(MFVi,j , i refers to the number of features and j means the number

of associated test cases, i.e., 10 i 1000 with an increment of 10
and 5 j 45 with an increment of 10) is dened to measure the
mean overall tness value for a certain number of runs nr as below
(in our case, nr = 100). OFVi,j,r is the tness value of the 25,000th
generation after the rth run with i features and j associated test
cases.
of the algorithms and different number of features and test cases

(Sheskin, 2007). The value of ranges from 1 to 1, i.e., there is a
positive correlation if is equal to 1 and a negative correlation when
is 1. A close to 0 shows that there is no correlation between the
two sets of data. Moreover, we also report signicance of correlation
using Prob > | |, a value lower than 0.05 means the correlation is
statistically signicant.
nr
r=1 OFVi,j,r
MFVi,j =
nr
Moreover, Mean Fitness Value for Features (MFVFi ) is dened

for RQ4 to measure how the increment of the number of features impacts the performance of the selected search algorithms as
below.
45
MFVFi =
j=5,j=j+10 MFVi,j
where 10 i 1000 with an increment of 10, i.e., a set of hundred

MFV_F values can be obtained for various numbers of features.
Similarly, Mean Fitness Value for associated Test Cases (MFVTCj ) is
also dened for RQ4 to measure how the increment of the number of
associated test cases impacts the performance of the selected multiobjective search algorithms as follows.
1000
MFVTCj =
i=10,i=i+10 MFVi,j
100
where 5 i 45 with an increment of 10, i.e., a set of ve MFV_TC

values can be obtained for different number of associated test cases.
5.2. Statistical tests

To compare the obtained result and given thresholds, the Vargha
and Delaney statistics, KruskalWallis test and MannWhitney U test
are used based on the guidelines for reporting statistical tests for
randomized algorithms (Arcuri and Briand, 2011). The Vargha and
Delaney statistics is used to calculate A12 , which is a non-parametric
effect size measure. In our context, A12 is used to compare the probability of yielding higher values for each objective function and overall
tness value (OFV) for two algorithms A and B. If A12 is 0.5, the two algorithms are equivalent. If A12 is greater than 0.5, A has higher chances
to obtain better solution than B. To determine statistical signicance
of results, rst the KruskalWallis test together with the Bonferroni
Correction is performed to determine if there are signicant differences among all the algorithms (Mcdonald, 2009). Notice that the
Bonferroni Correction is used for adjusting p-value obtained by the
KruskalWallis test since the KruskalWallis test is used to compare
multiple sets of results (samples) obtained by various search algorithms in our case. If signicant differences are observed, we further
perform pair-wise comparisons of the algorithms using the Mann
Whitney U test (Mcdonald, 2009). Therefore, when comparing the
performance of the selected search algorithms (RQ2 and RQ3), we
rst perform the KruskalWallis test to assess whether there exists
signicant differences for all the samples obtained by various algorithms. If signicant differences are observed, the MannWhitney U
test is further performed together with Vargha and Delaney statistics
for pair-wise comparison of the algorithms. We choose the signicance level of 0.05, i.e., there is a signicant difference if p-value is
less than 0.05. Based on the above description, we follow the following convention in the paper: An algorithm A has better performance than algorithm B, if the A12 value is greater than 0.5 and
have signicantly better performance than B if p-value is less than
0.05.
Moreover, to address RQ4, we choose the Spearmans rank correlation coecient ( ) to measure the relations between the MFV
5.3. Experiment execution

According to the guidelines in Arcuri and Briand (2011), each algorithm must be run for at least 100 times to account for random
variation inherited in search algorithms. All the selected algorithms
are run up to 25,000 generations each time and we collected the
optimal solution including the nal value of tness function for the
25,000th generation. We ran our experiments on two PCs with Intel
Core i7 2.3 GHz with 4 GB of RAM, running Microsoft Windows 7
operating system. In summary, it took 204.5 h (eight and half days)
to run the algorithms for all the problems.
6. Results and analysis
In this section, we present the results and analysis of the experiments for the individual research question including two parts: (1)
comparing the performance of the three weight-based GAs since all
of them are based on weight theory; and (2) comparing the performance of the best weight-based GA and the other selected search
algorithms, which are classied into different mechanisms to assess
which one can achieve the best performance for our minimization
problem. Recall that RQ1RQ3 address questions how the selected
algorithms perform in terms of cost/effectiveness for our minimization problem (as compared with RS), while RQ4 tackles the question corresponded with the impact on the performance of these
algorithms with the growth of features and associated test cases
(Section 5.1.1).
6.1. Results and analysis for the selected weight-based GAs
In this section, we present the results and analysis for the selected weight-based GAs (i.e., WBGA_W1 , WBGA_W2 , WBGA-MO
and RWGA) using our industrial case study and 500 articial problems. Furthermore, we also report a timing analysis, where we compare time taken by all the weight-based GAs as compared with
RS.
6.1.1. Results and analysis for industrial case study
We rst discuss the results and analysis for the industrial case
study for addressing the proposed research questions.
6.1.1.1. Results and analysis for research question 1. Table 4 shows
the results, when the performance of the selected weight-based algorithms (WBGA_W1 , WBGA_W2 , WBGA-MO and RWGA) are compared with the threshold values for each objective (i.e., TMP, FPC,
FDC, OET) for our industrial case study, while Table 5 describes the
results for the OFV (i.e., overall tness value) obtained by the selected algorithms. In these two tables, we did one sample Mann
Whitney test since we compared the results by different algorithms
with one set of xed values for TMP, FPC, FDC, OET and OFV. Notice that RS is a baseline for assessing the diculty of the addressed
problems.
Based on the obtained results, we can answer RQ1 as follows.
For all the objectives, WBGA-MO and RWGA have higher probability to obtain better results when compared with the given
thresholds, i.e., they have higher probability to be used when required since all of the A12 values are greater than 0.5 for TMP, FPC
ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
Table 4
Results for comparing the weight-based algorithms with the given thresholds for each objective.
Products
Algorithms
Objectives with the given thresholds

TMP: 0.8
FPC: 0.8
FDC: 0.85
OET: 2 h
C20
WBGA_W1
WBGA_W2
WGBA-MO
RWGA
RS
0.45
0.55
0.61
0.69
0.22
0.12
0.10
0.17
0.19
0.01
0.37
0.44
0.67
0.76
0.34
0.08
0.17
0.13
0.19
0.02
0.39
0.56
0.54
0.75
0.47
0.19
0.36
0.32
0.10
0.05
0.45
0.38
0.39
0.32
0.55
0.16
0.09
0.08
0.03
0.23
C40
WBGA_W1
WBGA_W2
WBGA-MO
RWGA
RS
0.55
0.68
0.57
0.68
0.35
0.40
0.04
0.46
0.08
0.02
0.44
0.39
0.63
0.65
0.23
0.18
0.20
0.11
0.10
0.01
0.41
0.56
0.61
0.73
0.44
0.16
0.08
0.10
0.05
0.11
0.39
0.42
0.38
0.34
0.65
0.08
0.19
0.07
0.06
0.03
C60
WBGA_W1
WBGA_W2
WBGA-MO
RWGA
RS
0.61
0.59
0.56
0.67
0.39
0.10
0.37
0.22
0.13
0.01
0.48
0.52
0.64
0.71
0.32
0.38
0.27
0.09
0.17
0.02
0.32
0.57
0.66
0.78
0.35
0.04
0.06
0.07
0.05
0.02
0.43
0.39
0.32
0.39
0.71
0.33
0.25
0.08
0.15
0.01
C90
WBGA_W1
WBGA_W2
WGBA-MO
RWGA
RS
0.64
0.66
0.67
0.70
0.22
0.21
0.18
0.19
0.12
0.02
0.30
0.53
0.58
0.64
0.34
0.04
0.12
0.19
0.17
0.01
0.27
0.37
0.64
0.69
0.49
0.03
0.10
0.12
0.09
0.22
0.41
0.43
0.39
0.31
0.67
0.27
0.35
0.10
0.06
0.02
A: A12 , p: p-value. All p-values less than 0.05 are identied as bold.
and FDC and all of the A12 values are less than 0.5 for OET. Moreover, there are no signicant differences (all p-values are greater
than 0.05) when comparing with the given thresholds as shown in
Table 4.
For WBGA_W1 and WBGA_W2 , the results do not stay stable.
For some products and objectives (e.g., for TMP in C40 product),
WBGA_W1 and WBGA_W2 have more chance to achieve better results than the given thresholds since the A12 values are greater than
0.5 (Table 4). But for the other products and objectives (e.g., for FDC in
C90 product), WBGA_W1 and WBGA_W2 have less chance to obtain
the better results as compared with the given thresholds since the
A12 values are much less than 0.5 (Table 4). Since the only difference
between WBGA_W1 and WBGA_W2 is the sets of weights, different
weights can result in total different results when using WBGA. So providing an appropriate set of weights before using WBGA is essential
to obtain the expected results.
Based on the results, we can see that RS always has less probability to be used in practice when compared with the given thresholds since all the A12 values are less than 0.5 for TMP, FPC and FDC
and all of the A12 values are greater than 0.5 for OET. Moreover,
there are signicant differences when comparing with the given
thresholds since most of the p-values are less than 0.05 as shown in
Table 4.
When comparing the results for OFV considering all the objectives
together, WGBA-MO and RWGA have higher probability to achieve
better OFV than the given thresholds since all the A12 values are
greater than 0.5 and there are almost no signicant differences (most
of the p-values are greater than 0.05) as shown in Table 5. As for
WBGA (WBGA_W1 and WBGA_W2 ), the results do not stay stable,

i.e., some of them have higher probability to obtain better OFV as
compared with the thresholds while some others have less probability to achieve better OFV than the thresholds (Table 5). However, RS
always has less probability to obtain better OFV than the thresholds
since all the A12 values are greater than 0.5 and all the p-values are
less than 0.05 (Table 5).
In summary, weight-based GAs can assist to achieve the given
thresholds. However, for WBGA (i.e., WGBA_W1 and WGBA_W2 ), determining appropriate weights is challenging as it mostly dependent
on domain-expertise and expert-guessing of what should be the most
appropriate values, which is the main reason why WBGA with various
weights can achieve different weights. But WGBA-MO and RWGA are
not dependent on the pre-dened weights thus the obtained results
are more stable.
6.1.1.2. Results and analysis for research question 2 and 3. For the
Saturn products, the KruskalWallis test (together with the Bonferroni Correction) was rst performed for all the samples obtained by
weight-based GAs and RS with respect to each objective and OFV
(considering all the objective together, Section 5.1.3) to determine
whether there are signicant differences between the performances
of the algorithms. We obtained the following p-values: 0.017 for
TMP, 0.008 for FPC, 0.020 for FDC, 0.012 for OET and 0.010 for OFV,
which shows that there exist signicant differences between the
performance of the weight-based GAs and RS since all the p-values
are less than 0.05. Therefore, the Vargha and Delaney statistics and
the MannWhitney U test were further performed to compare each
Table 5
Results for comparing the weight-based algorithms with the given thresholds for OFV.
Pair of algorithms
WBGA_W1 vs. Thresholds

WBGA_W2 vs. Thresholds
WBGA-MO vs. Thresholds
RWGA vs. Thresholds
RS vs. Thresholds
C20
C40
C60
C90
0.51
0.52
0.56
0.62
0.28
0.23
0.18
0.11
0.32
0.02
0.32
0.41
0.55
0.59
0.33
0.04
0.12
0.15
0.18
0.02
0.35
0.59
0.62
0.67
0.39
0.04
0.14
0.09
0.07
0.01
0.40
0.47
0.56
0.71
0.27
0.12
0.19
0.16
0.04
0.01
All p-values less than 0.05 are identied as bold.
ARTICLE IN PRESS
JID: JSS
10
[m5G;December 8, 2014;18:52]
Table 6
Results for comparing the weight-based algorithms for each objective.
Products
Pair of algorithms
Objectives
TMP
FPC
FDC
OET
C20
WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA-MO
WBGA_W1 vs. RWGA
WBGA_W1 vs. RS
WBGA_W2 vs. WBGA-MO
WBGA_W2 vs. RWGA
WBGA_W2 vs. RS
WBGA-MO vs. RWGA
WBGA-MO vs. RS
RWGA vs. RS
0.57
0.35
0.22
0.66
0.54
0.28
0.65
0.30
0.67
0.61
0.29
0.05
0.02
0.04
0.12
0.02
0.02
0.01
0.02
0.06
0.58
0.41
0.40
0.58
0.47
0.32
0.66
0.32
0.70
0.73
0.15
0.38
0.25
0.17
0.53
0.02
0.02
0.02
0.01
0.01
0.45
0.40
0.27
0.69
0.40
0.41
0.57
0.34
0.58
0.59
0.27
0.27
0.10
0.04
0.14
0.13
0.14
0.02
0.09
0.10
0.54
0.32
0.38
0.68
0.59
0.30
0.64
0.38
0.69
0.79
0.22
0.08
0.24
0.02
0.08
0.02
0.03
0.06
0.02
0.01
C40
WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA-MO
WBGA_W1 vs. RWGA
WBGA_W1 vs. RS
WBGA_W2 vs. WBGA-MO
WBGA_W2 vs. RWGA
WBGA_W2 vs. RS
WBGA-MO vs. RWGA
WBGA-MO vs. RS
RWGA vs. RS
0.52
0.45
0.18
0.67
0.48
0.31
0.61
0.33
0.66
0.82
0.44
0.38
0.01
0.03
0.21
0.03
0.03
0.02
0.03
0.01
0.47
0.35
0.26
0.60
0.42
0.40
0.59
0.29
0.69
0.76
0.16
0.29
0.09
0.18
0.23
0.09
0.04
0.02
0.02
0.02
0.51
0.38
0.41
0.69
0.52
0.29
0.65
0.37
0.60
0.69
0.45
0.22
0.37
0.03
0.37
0.02
0.03
0.03
0.04
0.02
0.56
0.33
0.32
0.70
0.43
0.37
0.67
0.39
0.65
0.59
0.30
0.14
0.11
0.01
0.12
0.10
0.02
0.08
0.03
0.01
C60
WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA-MO
WBGA_W1 vs. RWGA
WBGA_W1 vs. RS
WBGA_W2 vs. WBGA-MO
WBGA_W2 vs. RWGA
WBGA_W2 vs. RS
WBGA-MO vs. RWGA
WBGA-MO vs. RS
RWGA vs. RS
0.53
0.52
0.35
0.59
0.57
0.30
0.60
0.39
0.60
0.59
0.43
0.46
0.04
0.27
0.09
0.03
0.04
0.14
0.03
0.03
0.58
0.36
0.17
0.65
0.43
0.21
0.59
0.29
0.68
0.62
0.15
0.19
0.01
0.04
0.22
0.01
0.04
0.02
0.02
0.02
0.46
0.29
0.25
0.68
0.47
0.29
0.61
0.30
0.55
0.74
0.21
0.02
0.07
0.04
0.39
0.01
0.03
0.03
0.16
0.01
0.47
0.34
0.34
0.56
0.54
0.32
0.56
0.32
0.64
0.66
0.34
0.11
0.15
0.29
0.27
0.02
0.11
0.03
0.02
0.03
C90
WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA-MO
WBGA_W1 vs. RWGA
WBGA_W1 vs. RS
WBGA_W2 vs. WBGA-MO
WBGA_W2 vs. RWGA
WBGA_W2 vs. RS
WBGA-MO vs. RWGA
WBGA-MO vs. RS
RWGA vs. RS
0.48
0.40
0.26
0.57
0.44
0.37
0.62
0.36
0.64
0.69
0.39
0.15
0.03
0.10
0.20
0.05
0.03
0.03
0.03
0.01
0.57
0.31
0.36
0.67
0.46
0.40
0.65
0.39
0.67
0.71
0.16
0.07
0.12
0.03
0.31
0.13
0.02
0.05
0.02
0.01
0.55
0.41
0.38
0.59
0.48
0.31
0.68
0.30
0.70
0.77
0.21
0.17
0.24
0.08
0.23
0.03
0.02
0.02
0.01
0.01
0.49
0.26
0.19
0.65
0.42
0.29
0.60
0.35
0.60
0.63
0.44
0.04
0.01
0.04
0.10
0.01
0.03
0.03
0.03
0.02
pair of weight-based GAs as shown in Tables 6 and 7. In total, we

have C25 = 10 pairs to compare for the three selected search algorithms. The obtained results are rst compared for each objective
(Table 6) and then compared for OFV considering all the objectives
(Table 7).
Based on the obtained results, we can answer RQ2 and RQ 3 as
follows:
RQ2: For all the objectives (i.e., TMP, FPC, FDC and OET), the results show that all the selected weight-based GAs have signicantly
better performance than RS since all the A12 values for TMP, FPC,
and FDC are greater than 0.5, and all the A12 values for OET are
less than 0.5, and most of the p-values are less than 0.05 (Table 6).
When comparing the OFV results considering all objectives for the
algorithms and RS, we can see that all of the selected weight-based
Table 7
Results for comparing the weight-based algorithms for OFV.
Pair of algorithms
WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA-MO
WBGA_W1 vs. RWGA
WBGA_W1 vs. RS
WBGA_W2 vs. WBGA-MO
WBGA_W2 vs. RWGA
WBGA_W2 vs. RS
WBGA-MO vs. RWGA
WBGA-MO vs. RS
RWGA vs. RS
C20
C40
C60
C90
0.41
0.36
0.18
0.60
0.39
0.33
0.60
0.39
0.60
0.66
0.23
0.29
0.01
0.04
0.04
0.02
0.04
0.04
0.03
0.01
0.42
0.32
0.27
0.69
0.40
0.30
0.59
0.32
0.69
0.79
0.38
0.05
0.03
0.01
0.12
0.02
0.04
0.02
0.02
0.02
0.51
0.40
0.21
0.57
0.37
0.38
0.63
0.37
0.62
0.65
0.46
0.26
0.02
0.11
0.04
0.04
0.03
0.04
0.03
0.02
0.46
0.29
0.22
0.63
0.43
0.31
0.67
0.32
0.64
0.72
0.26
0.02
0.02
0.03
0.22
0.02
0.03
0.02
0.03
0.02
ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
Table 8
Results for the Vargha and Delaney statistics-articial problems (without/with Whitney U test).
RQ
Pair of algorithms (A vs. B)
A>B
A<B
A=B
RQ2
WBGA_W1 vs. RS
WBGA_W2 vs. RS
WBGA_MO vs. RS
RWGA vs. RS
367/354
392/371
445/426
489/480
133/102
108/96
55/45
11/0
0/54
0/33
0/29
0/20
RQ3
WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA-MO
WBGA_W1 vs. RWGA
WBGA_W2 vs. WBGA-MO
WBGA_W2 vs. RWGA
WBGA-MO vs. RWGA
226/202
198/170
119/93
211/198
134/122
192/174
274/264
302/287
381/371
289/264
366/354
308/295
0/34
0/43
0/36
0/38
0/24
0/31
GAs have signicantly better performance than RS since all the A12
values are greater than 0.5 and all the p-values are less than 0.05
(Table 7).
RQ3: For all the objectives (i.e., TMP, FPC, FDC and OET), the results
show that RWGA signicantly outperforms WBGA_W1 , WBGA_W2
and WBGA-MO since all the A12 values for TMP, FPC, and FDC are
greater than 0.5, and all the A12 values for OET are less than 0.5,
and most of the p-values are less than 0.05 (Table 6). Similarly,
when comparing OFV, the performance of RWGA is also signicantly better than the other weight-based GAs since all the A12
values are greater than 0.5 and all the p-values are less than 0.05
(Table 7).
In general, we can conclude that RWGA achieves the best performance among the selected weight-based GAs for all the objectives
and overall tness value (OFV) when considering all the objectives
together.
6.1.2. Results and analysis for articial problems
To answer RQ2 and RQ3, we rst conducted the KruskalWallis
test (together with the Bonferroni Correction) for all the obtained results and we obtained a p-value of 0.014 (less than 0.05) showing that
there are signicant differences between the results obtained by the
selected weight-based GAs and RS. Moreover, we further compared
the selected weight-based algorithms with RS and then compared
each pair of them based on mean tness value (MFV) obtained after 25,000 generations for each algorithm and each of the problems.
Recall that each problem was repeated for 100 times to account for
random variation (i.e., MFV is a mean for 500 values of OFV after 100
runs). Notice that we answer RQ4 in terms of scalablity for all the ten
search algorithms in Section 6.2.2.2.
Table 8 summarizes the results for Vargha and Delaney statistics
without/with Whitney U test for the 500 articial problems. Two
numbers are shown in each cell of the table splited by a slash. The
rst number in the column A > B shows the number of problems
out of 500 for which an algorithm A has better performance than B
(A12 > 0.5), A < B means vice versa (A12 < 0.5), and A = B means the
number of problems for which A has equivalent performance as B
(A12 = 0.5). The second number after / in the column A > B means
the number of problems out of 500 for which an algorithm A has
signicantly better performance than B (A12 > 0.5 and p < 0.05), A
< B means vice versa (A12 < 0.5 and p < 0.05), and A = B means the
number of problems for which there are no siginicant differences in
performance between A and B (p 0.05).
6.1.2.1. Results and analysis for research question 2 and 3. To answer RQ2 and RQ3, we compared WBGA_W1 , WBGA_W2 , WBGA-MO,
RWGA and RS based on the mean tness values for each problem
(MFVi,j , where 10 i 1000 with an increment of 10 and 5 j 45
with an increment of 10, i.e., there are 500 MFV values for the 500
articial problems).
11
Table 9
Average running time for the selected weight-based search algorithms.
Algorithms
WBGA_W1
WBGA_W2
WBGA-MO
RWGA
RS
Time (s)
2.01
2.11
1.94
1.37
1.13
RQ2: All the selected weight-based GAs signicantly outperformed RS for nding optimal solutions in most of the problems.
For instance, WBGA-MO has better performance than RS for 445
problems and the performance is signicantly better for 426 problems. WBGA-MO has worse performance than RS for 55 problems (45 of them were signicantly worse). There were no signicant differences between WBGA-MO and RS for 29 problems
(Table 8).
RQ3: RWGA achieved the best performance among all the other
algorithms in most of the problems (the performance of RWGA
is better than the other algorithms for 351.7 problems on average ((381 + 366 + 308)/3 = 351.7) and signicantly better for 340
problems on average ((371 + 354 + 295)/3 = 340)). WBGA_W1 and
WBGA_W2 have almost equivalent performance, whose performance
stays at the second level. Moreover, WBGA-MO achieved better
performance than WBGA_W1 and WBGA_W2 (WBGA-MO outperformed WBGA_W1 and WBGA_W2 for 302 and 289 problems respectively and signicantly outperformed for 287 and 264 problems
respectively).
6.1.3. Timing analysis for weight-based GAs
In addition, we report the average time taken by the three weightbased GAs per run (i.e., WBGA, WBGA-MO and RWGA) as compared with RS (shown in Table 9). In addition, the KruskalWallis
test was conducted to evaluate whether there are signicant differences among the running time of the weight-based GAs as compared with RS (100 run in our case). The p-value obtained by the
KruskalWallis test is 0.39, which shows that there are no signicant differences among the running time of the weight-based GAs
and RS. Based on the results, we can conclude that the selected
weight-based GAs took similar running time as compared with RS
in terms of nding an optimal solution for our test minimization
problem.
6.2. Results and analysis for the best weight-based GA (RWGA) and the
other selected search algorithms
This section presents the results and related analysis for our industrial case study and the designed articial problems using the best
weight-based GA (RWGA) and the other selected algorithms.
6.2.1. Results and analysis for industrial case study
In this section, we summarize the results in the following sections
for the industrial case study.
6.2.1.1. Analysis for research question 1. We did one sample Mann
Whitney test since we compared the results by different algorithms
with one set of xed values for TMP, FPC, FDC, OET and OFV (i.e.,
overall tness value). Tables 10 and 11 show that the performance of
algorithms as compared with the threshold values for each objective
(i.e., TMP, FPC, FDC, OET and OFV). Based on the results, we answer
RQ1 as follows.
For TMP, all the selected algorithms except RS have higher probability to achieve better results than the given threshold for TMP since
most of the A12 values are greater than 0.5 and there are no signicant
differences between the results and the threshold (all the p-values
are greater than 0.05). However, all the values of A12 obtained by RS
are less than 0.5 and all the p-values are less than 0.05, i.e., RS has
ARTICLE IN PRESS
JID: JSS
12
[m5G;December 8, 2014;18:52]
Table 10
Results for comparing the eight algorithms with the given thresholds for each objective.
Products
Algorithms
Objectives with the given thresholds

TMP: 0.8
FPC: 0.8
FDC: 0.85
OET: 2 h
C20
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
0.69
0.72
0.65
0.59
0.68
0.57
0.63
0.22
0.19
0.23
0.32
0.18
0.24
0.32
0.35
0.01
0.76
0.63
0.61
0.38
0.54
0.62
0.68
0.34
0.19
0.33
0.34
0.04
0.27
0.31
0.34
0.02
0.75
0.73
0.57
0.39
0.54
0.63
0.57
0.47
0.10
0.21
0.17
0.21
0.18
0.35
0.29
0.05
0.32
0.29
0.34
0.25
0.43
0.39
0.28
0.55
0.03
0.01
0.03
0.02
0.12
0.08
0.02
0.23
C40
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
0.68
0.55
0.72
0.57
0.61
0.65
0.72
0.35
0.08
0.22
0.31
0.25
0.34
0.17
0.19
0.02
0.65
0.65
0.58
0.31
0.62
0.57
0.61
0.23
0.10
0.18
0.41
0.04
0.19
0.23
0.31
0.01
0.73
0.68
0.61
0.33
0.54
0.64
0.59
0.44
0.05
0.32
0.24
0.04
0.19
0.27
0.34
0.11
0.34
0.34
0.27
0.25
0.42
0.38
0.26
0.65
0.06
0.02
0.04
0.02
0.17
0.07
0.02
0.03
C60
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
0.67
0.61
0.54
0.65
0.47
0.49
0.54
0.39
0.13
0.32
0.22
0.34
0.41
0.25
0.18
0.01
0.71
0.55
0.63
0.49
0.57
0.63
0.59
0.32
0.17
0.23
0.18
0.17
0.23
0.29
0.32
0.02
0.78
0.62
0.57
0.31
0.62
0.55
0.59
0.35
0.05
0.11
0.24
0.03
0.18
0.23
0.31
0.02
0.39
0.48
0.35
0.37
0.42
0.30
0.29
0.71
0.15
0.21
0.04
0.03
0.11
0.04
0.03
0.01
C90
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
0.70
0.58
0.61
0.48
0.54
0.63
0.68
0.22
0.12
0.25
0.27
0.14
0.11
0.35
0.29
0.02
0.64
0.69
0.54
0.47
0.49
0.55
0.59
0.34
0.17
0.17
0.21
0.09
0.15
0.23
0.17
0.01
0.69
0.57
0.62
0.43
0.54
0.65
0.71
0.49
0.09
0.34
0.17
0.08
0.13
0.25
0.23
0.22
0.31
0.28
0.34
0.24
0.42
0.31
0.23
0.67
0.06
0.03
0.07
0.02
0.13
0.04
0.02
0.02
signicantly less probability to obtain better results than the given

threshold (Table 10).
As for FPC and FDC, all the selected algorithms except SPEA2
and RS have higher chances to obtain better results than the given
thresholds when required since almost all the values of A12 are
greater than 0.5 and there are no signicant differences (all the
p-values are greater than 0.05). For SPEA2, all the A12 values are
less than 0.5 and half of the p-values are less than 0.05, i.e., SPEA2
has signicantly less probability to yield better results than the
given thresholds. RS also has signicantly less chance to be used
for achieving an acceptable level of FPC and FDC since all the values of A12 are less than 0.5 and most of the p-values are less than 0.05
(Table 10).
For OET, most of the A12 values achieved by the selected algorithms except RS are less than 0.5 and more than half of the
p-values are less than 0.05, i.e., they have signicantly higher chances
to meet the time budget within the given threshold (2 h). Moreover, all the A12 values obtained by RS are greater than 0.5 and
most of the p-values are less than 0.05, i.e., the whole execution time
achieved by RS has higher chance to be signicantly more than 2 h
(Table 10).
When comparing results for OFV considering all the objectives, the
selected algorithms except SPEA2 and RS have higher probability to
obtain better OFV when compared with the given thresholds since
all the A12 values are greater than 0.5 and there are no signicant
differences (all the p-values are greater than 0.05). As for SPEA2, the
obtained OFV has signicantly less chance to be equivalent as the
given thresholds (most of the A12 values are less than 0.5 and half
of the p-values are less than 0.05). For RS, the obtained OFV has also
higher chance to be signicantly less than the thresholds since all the
Table 11
Results for comparing the eight algorithms with the given thresholds for OFV.
Pair of algorithms
RWGA vs. Thresholds

NSGA-II vs. Thresholds
MOCell vs. Thresholds
SPEA2 vs. Thresholds
PAES vs. Thresholds
SMPSO vs. Thresholds
CellDE vs. Thresholds
RS vs. Thresholds
C20
C40
C60
C90
0.62
0.71
0.58
0.38
0.55
0.65
0.59
0.28
0.32
0.28
0.24
0.11
0.23
0.27
0.18
0.02
0.59
0.62
0.62
0.31
0.58
0.72
0.63
0.33
0.18
0.21
0.19
0.04
0.16
0.34
0.22
0.02
0.67
0.54
0.64
0.29
0.61
0.69
0.74
0.39
0.07
0.12
0.21
0.03
0.27
0.27
0.39
0.01
0.71
0.66
0.59
0.51
0.60
0.74
0.69
0.27
0.04
0.19
0.16
0.13
0.33
0.25
0.35
0.01
ARTICLE IN PRESS
JID: JSS
Table 12
Number of solutions obtained by each search algorithm.
Algorithms
WBGA_W1
WBGA_W2
WBGA-MO
RWGA
NSGA-II
MOCell
Solution No.
11
15
SPEA2
PAES
SMPSO
CellDE
RS
19
20
A12 values are less than 0.5 and all the p-values are less than 0.05
(Table 11).
We also report the number of solutions that satisfy all the given
thresholds (i.e., 0.8 for TMP, 0.8 for FPC, 0.85 for FDC and 2 h for OET)
obtained by each selected algorithm as shown in Table 12. Based on
the results, we observe that weight-based GAs can manage to nd
one solution which balances all the objectives together (i.e., maximum of the OFV value) while the Pareto-based algorithms can obtain
a set of non-dominated solutions, e.g., eleven non-dominiated solutions are obtained by using NSGA-II. Notice RS can not manage to
nd a solution which can satisfy the required thresholds for all the
objectives.
In general, the selected algorithms except SPEA2 and RS can
achieve the given thresholds for each objective well and the overall performance still works well in terms of OFV.
6.2.1.2. Analysis for research question 2 and 3. Similarly as Section
6.1.1.2, for the Saturn products, all the p-values obtained by the
KruskalWallis test (together with the Bonferroni Correction) are
less than 0.05 for each objective and OFV (0.009 for TMP, 0.031 for
FPC, 0.015 for FDC, 0.024 for OET and 0.016 for OFV), showing that
there are signicant differences between the performance of the selected search algorithms for each objective and OFV. Therefore, we
further performed the Vargha and Delaney statistics and the Mann
Whitney U test to compare each pair of the eight search algorithms
as shown in Tables 13 and 14. In total, we have C28 = 28 pairs to
compare for the eight selected search algorithms. The obtained results are rst compared for each objective and then compared for
OFV considering all the objectives (shown as Tables 13 and 14, respectively). Based on the results, we answer the RQ2 and RQ3 as
follows.
RQ2: For TMP, FPC, FDC and OET, we rst compared the performance of the selected algorithms with RS. The results show that
all the selected search algorithms have signicantly better performance than RS for TMP and OET since all the A12 values for TMP
are greater than 0.5, and all the A12 values for OET are less than
0.5, and most of the p-values are less than 0.05. For FPC and FDC,
all the selected algorithms except SPEA2 achieve signicantly better performance than RS since almost all the A12 values are greater
than 0.5 and most of the p-values are less than 0.05. SPEA2 has almost equivalent performance as RS for FPC and FDC since most of
the A12 values are close to 0.5 and the p-values are greater than 0.05
(Table 13).
When comparing the OFV results considering all objectives for the
algorithms and RS, we can see that six of them (i.e., RWGA, NSGA-II,
MOCell, PAES, SMPSO and CellDE) have signicantly better performance than RS since all the A12 values are greater than 0.5 and all the
p-values are less than 0.05. SPEA2 has almost equivalent performance
as RS since most of the A12 values are closer to 0.5 and the p-values
are greater than 0.05 (Table 14).
RQ3: For TMP, CellDE achieved the best performance than the other
algorithms since most of the A12 values are greater than 0.5 and the pvalues are less than 0.05. Similarly, MOCell, NSGA-II and SPEA2 have
the best performance for FPC, FDC and OET among all the algorithms,
respectively (Table 13).
When comparing the OFV results considering all objectives together, RWGA performs signicantly better among all the search
[m5G;December 8, 2014;18:52]
13
algorithms since all the A12 values are greater than 0.5 and most
of p-values are less than 0.05. Moreover, SPEA2 has signicantly
worse performance than the other six search algorithms since most
of the A12 values are less than 0.5 and p-values are greater than 0.05
(Table 14).
In general, we can conclude that different algorithms can achieve
the best performance for different objectives (e.g., CellDE for the objective TMP) but RWGA has signicantly better performance than the
other search algorithms when considering all the objectives together
in general.
6.2.2. Results and analysis for articial problems
In this section, we summarize the results as the following sections
for the designed 500 articial problems.
6.2.2.1. Results and analysis for research question 2 and 3. To answer RQ2 and RQ3, the KruskalWallis test (together with the Bonferroni Correction) was rst performed for the results of the eight
search algorithms and we obtained a p-value of 0.024 (less than 0.05)
showing that there exist signicant differences between the performances of the selected algorithms. Therefore, we further compared
the algorithms (i.e., RWGA, NSGA-II, MOCell, SPEA2, PAES, SMPSO,
CellDE) with RS and then compared each pair of them based on mean
tness value (MFV) achieved after 25,000 generations for each algorithm and each of the 500 problems. Recall that each problem
was repeated for 100 times to account for random variation, i.e.,
MFV is a mean for 100 values of OFV after 100 runs (Section 5.1.3).
Table 14 summarizes the results for Vargha and Delaney statistics
without or with Whitney U test for the 500 articial problems (similarly as Table 13).
RQ2: RWGA, NSGA-II, MOCell, PAES, SMPSO, and CellDE achieved
signicantly better performance than RS for nding optimal solutions for our minimization problem in most of the problems. For
instance, RWGA has better performance than RS for 489 problems
and the performance is signicantly better for 480 problems. RWGA
has worse performance than RS for 11 problems (none of them
was signicantly worse). There were no signicant differences between RWGA and RS for 20 problems. The performance of SPEA2
is close to RS (SPEA2 is better than RS for 298 problems and signicantly better for 251 problems. SPEA2 is worse than RS for 202
problems and signicantly worse for 179 problems. Meanwhile,
there is no signicant difference between them for 70 problems)
(Table 15).
RQ3: RWGA achieved the best performance among all the
other six algorithms in most of the problems, i.e., the performance of RWGA is better than the other six algorithms for
average 418.8 problems ((432 + 403 + 464 + 381 + 453 + 380)/6 =
418.8) and signicantly better for 401.7 problems on average
((405 + 397 + 451 + 362 + 438 + 357)/6 = 401.7). NSGA-II, MOCell,
PAES, SMPSO, and CellDE have almost equivalent performance,
whose performance stay at the second level. Finally, the performance of SPEA2 is worse than the above-mentioned algorithms, i.e.,
SPEA2 achieved the worst performance in the 500 articial problems (SPEA2 is worse than the other search algorithms for 377.8
problems and signicantly worse for 354.7 problems on average)
(Table 15).
6.2.2.2. Results and analysis for research question 4. To answer RQ4,
we calculated Spearmans rank correlation bettween mean tness values for each algorithm with increasing number of features (MFV_Fi ,
where 10 i 1000 with an increment of 10, Section 5.1.3) and associated test cases (MFV_TCj , where 5 j 45 with an increment of
10, Section 5.1.3). In this case, a xed number of features is used
to represent a product and each feature is associated with different
number of test cases. So we have 100 5 = 500 articial problems
ARTICLE IN PRESS
JID: JSS
14
[m5G;December 8, 2014;18:52]
Table 13
Results for comparing the eight algorithms for each objective.
Pair of algorithms
Objectives
TMP: 0.8
FPC: 0.8
FDC: 0.85
OET: 2 h
(1) Product C20

RWGA vs. NSGA-II
RWGA vs. MOCell
RWGA vs. SPEA2
RWGA vs. PAES
RWGA vs. SMPSO
RWGA vs. CellDE
RWGA vs. RS
NSGA-II vs. MoCell
NSGA-II vs. SPEA2
NSGA-II vs. PAES
NSGA-II vs. SMPSO
NSGA-II vs. CellDE
NSGA-II vs. RS
MOCell vs. SPEA2
MOCell vs. PAES
MOCell vs. SMPSO
MOCell vs. CellDE
MOCell vs. RS
SPEA2 vs. PAES
SPEA2 vs. SMPSO
SPEA2 vs. CellDE
SPEA2 vs. RS
PAES vs. SMPSO
PAES vs. CellDE
PAES vs. RS
SMPSO vs. CellDE
SMPSO vs. RS
CellDE vs. RS
0.43
0.51
0.48
0.47
0.42
0.36
0.61
0.51
0.46
0.54
0.46
0.49
0.72
0.58
0.56
0.54
0.31
0.69
0.51
0.45
0.28
0.69
0.43
0.46
0.67
0.31
0.65
0.76
0.11
0.12
0.11
0.21
0.24
0.04
0.06
0.17
0.23
0.28
0.17
0.35
0.01
0.17
0.14
0.27
0.03
0.04
0.34
0.22
0.01
0.02
0.29
0.31
0.02
0.01
0.03
0.01
0.34
0.37
0.57
0.45
0.52
0.43
0.73
0.41
0.75
0.53
0.55
0.47
0.82
0.75
0.68
0.65
0.66
0.79
0.34
0.29
0.31
0.52
0.46
0.43
0.63
0.55
0.69
0.67
0.04
0.04
0.07
0.18
0.13
0.08
0.01
0.13
0.01
0.24
0.18
0.23
0.02
0.01
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.21
0.17
0.25
0.03
0.24
0.03
0.03
0.40
0.36
0.29
0.54
0.47
0.52
0.59
0.67
0.70
0.65
0.68
0.64
0.73
0.68
0.54
0.49
0.51
0.69
0.36
0.31
0.28
0.54
0.49
0.47
0.68
0.53
0.72
0.77
0.08
0.04
0.03
0.25
0.28
0.14
0.10
0.02
0.02
0.02
0.02
0.02
0.01
0.02
0.19
0.25
0.31
0.02
0.01
0.01
0.01
0.17
0.24
0.29
0.02
0.19
0.01
0.01
0.53
0.55
0.37
0.51
0.53
0.55
0.79
0.48
0.37
0.53
0.48
0.54
0.65
0.35
0.52
0.46
0.43
0.67
0.71
0.67
0.72
0.75
0.48
0.45
0.67
0.51
0.66
0.63
0.25
0.19
0.03
0.22
0.17
0.21
0.01
0.24
0.02
0.25
0.17
0.28
0.01
0.01
0.14
0.21
0.27
0.02
0.01
0.02
0.01
0.01
0.17
0.23
0.02
0.34
0.02
0.02
(2) Product C40

RWGA vs. NSGA-II
RWGA vs. MOCell
RWGA vs. SPEA2
RWGA vs. PAES
RWGA vs. SMPSO
RWGA vs. CellDE
RWGA vs. RS
NSGA-II vs. MoCell
NSGA-II vs. SPEA2
NSGA-II vs. PAES
NSGA-II vs. SMPSO
NSGA-II vs. CellDE
NSGA-II vs. RS
MOCell vs. SPEA2
MOCell vs. PAES
MOCell vs. SMPSO
MOCell vs. CellDE
MOCell vs. RS
SPEA2 vs. PAES
SPEA2 vs. SMPSO
SPEA2 vs. CellDE
SPEA2 vs. RS
PAES vs. SMPSO
PAES vs. CellDE
PAES vs. RS
SMPSO vs. CellDE
SMPSO vs. RS
CellDE vs. RS
0.48
0.49
0.52
0.49
0.47
0.32
0.82
0.48
0.51
0.52
0.49
0.29
0.89
0.54
0.49
0.51
0.29
0.73
0.47
0.49
0.32
0.73
0.47
0.44
0.75
0.28
0.69
0.76
0.21
0.12
0.23
0.16
0.19
0.02
0.01
0.20
0.11
0.16
0.20
0.02
0.02
0.19
0.09
0.23
0.01
0.02
0.22
0.31
0.01
0.01
0.17
0.12
0.01
0.01
0.02
0.01
0.46
0.41
0.52
0.53
0.48
0.47
0.76
0.43
0.73
0.55
0.52
0.49
0.81
0.77
0.73
0.69
0.68
0.76
0.31
0.36
0.28
0.55
0.52
0.49
0.74
0.51
0.73
0.74
0.17
0.07
0.18
0.20
0.27
0.14
0.02
0.16
0.01
0.18
0.20
0.09
0.01
0.01
0.01
0.02
0.03
0.01
0.01
0.02
0.01
0.13
0.23
0.31
0.01
0.36
0.01
0.01
0.38
0.32
0.31
0.52
0.45
0.55
0.69
0.69
0.73
0.69
0.64
0.72
0.77
0.62
0.56
0.53
0.54
0.73
0.29
0.36
0.35
0.47
0.54
0.54
0.66
0.48
0.77
0.70
0.03
0.03
0.03
0.21
0.29
0.20
0.02
0.02
0.01
0.02
0.03
0.01
0.01
0.06
0.09
0.18
0.27
0.01
0.01
0.04
0.02
0.22
0.13
0.22
0.02
0.27
0.01
0.01
0.34
0.51
0.33
0.54
0.49
0.47
0.59
0.51
0.33
0.48
0.49
0.52
0.75
0.32
0.48
0.49
0.47
0.66
0.77
0.59
0.68
0.81
0.52
0.47
0.61
0.53
0.70
0.69
0.02
0.22
0.03
0.19
0.23
0.18
0.01
0.17
0.01
0.21
0.08
0.14
0.01
0.01
0.22
0.32
0.18
0.02
0.01
0.07
0.01
0.01
0.25
0.18
0.05
0.24
0.01
0.02
(3) Product C60

RWGA vs. NSGA-II
RWGA vs. MOCell
RWGA vs. SPEA2
RWGA vs. PAES
RWGA vs. SMPSO
RWGA vs. CellDE
RWGA vs. RS
NSGA-II vs. MoCell
NSGA-II vs. SPEA2
NSGA-II vs. PAES
NSGA-II vs. SMPSO
NSGA-II vs. CellDE
0.52
0.53
0.48
0.51
0.46
0.34
0.59
0.54
0.53
0.57
0.46
0.32
0.23
0.19
0.30
0.31
0.17
0.02
0.03
0.23
0.14
0.09
0.15
0.05
0.46
0.48
0.47
0.56
0.49
0.49
0.62
0.49
0.75
0.51
0.57
0.53
0.17
0.39
0.14
0.13
0.32
0.37
0.02
0.32
0.01
0.33
0.11
0.16
0.33
0.28
0.27
0.57
0.52
0.56
0.74
0.63
0.73
0.74
0.69
0.75
0.02
0.01
0.01
0.11
0.27
0.13
0.01
0.04
0.01
0.01
0.02
0.01
0.55
0.54
0.41
0.46
0.46
0.44
0.66
0.54
0.39
0.57
0.52
0.48
0.24
0.18
0.06
0.16
0.13
0.22
0.03
0.22
0.03
0.11
0.09
0.24
(continued on next page)
ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
15
Table 13 (continued)
Pair of algorithms
Objectives
TMP: 0.8
FPC: 0.8
FDC: 0.85
OET: 2 h
NSGA-II vs. RS
MOCell vs. SPEA2
MOCell vs. PAES
MOCell vs. SMPSO
MOCell vs. CellDE
MOCell vs. RS
SPEA2 vs. PAES
SPEA2 vs. SMPSO
SPEA2 vs. CellDE
SPEA2 vs. RS
PAES vs. SMPSO
PAES vs. CellDE
PAES vs. RS
SMPSO vs. CellDE
SMPSO vs. RS
CellDE vs. RS
0.57
0.54
0.48
0.51
0.40
0.74
0.55
0.49
0.34
0.63
0.48
0.20
0.73
0.27
0.70
0.73
0.03
0.22
0.24
0.29
0.05
0.01
0.17
0.35
0.02
0.05
0.27
0.01
0.01
0.01
0.01
0.01
0.65
0.67
0.72
0.73
0.69
0.77
0.36
0.26
0.28
0.54
0.53
0.47
0.76
0.53
0.68
0.75
0.01
0.02
0.01
0.01
0.02
0.01
0.02
0.01
0.01
0.12
0.19
0.31
0.01
0.29
0.02
0.01
0.81
0.72
0.51
0.53
0.54
0.70
0.35
0.29
0.24
0.52
0.44
0.53
0.70
0.48
0.75
0.79
0.02
0.01
0.30
0.21
0.23
0.01
0.02
0.01
0.01
0.29
0.16
0.14
0.01
0.20
0.01
0.01
0.59
0.27
0.56
0.53
0.49
0.74
0.65
0.73
0.67
0.78
0.53
0.48
0.75
0.55
0.74
0.80
0.04
0.01
0.09
0.18
0.35
0.01
0.03
0.01
0.01
0.02
0.20
0.33
0.01
0.17
0.01
0.01
(4) Product C90

RWGA vs. NSGA-II
RWGA vs. MOCell
RWGA vs. SPEA2
RWGA vs. PAES
RWGA vs. SMPSO
RWGA vs. CellDE
RWGA vs. RS
NSGA-II vs. MoCell
NSGA-II vs. SPEA2
NSGA-II vs. PAES
NSGA-II vs. SMPSO
NSGA-II vs. CellDE
NSGA-II vs. RS
MOCell vs. SPEA2
MOCell vs. PAES
MOCell vs. SMPSO
MOCell vs. CellDE
MOCell vs. RS
SPEA2 vs. PAES
SPEA2 vs. SMPSO
SPEA2 vs. CellDE
SPEA2 vs. RS
PAES vs. SMPSO
PAES vs. CellDE
PAES vs. RS
SMPSO vs. CellDE
SMPSO vs. RS
CellDE vs. RS
0.22
0.53
0.55
0.52
0.44
0.38
0.69
0.53
0.49
0.57
0.49
0.38
0.91
0.57
0.52
0.55
0.31
0.68
0.54
0.52
0.26
0.75
0.20
0.54
0.68
0.32
0.73
0.78
0.01
0.16
0.14
0.19
0.11
0.02
0.01
0.15
0.37
0.14
0.33
0.06
0.02
0.12
0.21
0.09
0.02
0.02
0.16
0.29
0.01
0.01
0.01
0.19
0.02
0.03
0.01
0.01
0.36
0.45
0.51
0.49
0.52
0.53
0.71
0.44
0.68
0.56
0.57
0.54
0.93
0.68
0.77
0.72
0.74
0.80
0.28
0.40
0.32
0.51
0.55
0.44
0.77
0.55
0.70
0.81
0.01
0.14
0.30
0.34
0.22
0.19
0.01
0.20
0.02
0.16
0.09
0.17
0.01
0.02
0.01
0.01
0.01
0.01
0.01
0.06
0.01
0.39
0.17
0.10
0.01
0.24
0.01
0.01
0.44
0.26
0.37
0.56
0.49
0.51
0.77
0.63
0.77
0.69
0.73
0.70
0.81
0.72
0.54
0.48
0.55
0.61
0.34
0.26
0.30
0.43
0.57
0.59
0.73
0.53
0.68
0.83
0.16
0.01
0.04
0.14
0.34
0.30
0.01
0.06
0.01
0.02
0.01
0.02
0.01
0.01
0.14
0.22
0.18
0.05
0.02
0.01
0.01
0.16
0.11
0.08
0.01
0.28
0.02
0.01
0.33
0.53
0.30
0.59
0.54
0.52
0.63
0.46
0.31
0.48
0.55
0.51
0.81
0.36
0.49
0.54
0.49
0.74
0.75
0.52
0.73
0.58
0.48
0.54
0.74
0.49
0.74
0.75
0.02
0.17
0.02
0.08
0.20
0.23
0.02
0.20
0.01
0.31
0.14
0.33
0.01
0.02
0.31
0.12
0.29
0.01
0.01
0.28
0.01
0.12
0.20
0.13
0.01
0.29
0.01
0.01
corresponding to minimizing the test suites for various products.

Table 16 shows the results of correlation between MFV_Fi and the
increasing number of features, and between MFV_TCj with the increasing associated test cases for various algorithms using Spearmans
rank correlation ( ). Recall that a higher value of MFV_Fi or MFV_TCj
represents a better performance of an algorithm.
Increasing Number of Features: For weight-based GAs (i.e.,
WBGA_W1 , WBGA_W2 , WBGA-MO and RWGA), we observed that
the values of MFV_F increased signicantly as the increasing number of features since the values for Spearmans are greater than 0
(0.52, 0.47, 0.69 and 0.74 respectively) and the p-value is less than
0.0001, i.e., the performance increases signicantly with the growth
of features number. For SPEA2, PAES, SMPSO and CellDE, the values
of MFV_F also increase but not signicantly as the increasing features
since Spearmans is greater than 0 and p-values are greater than
0.05, i.e., the performances of SPEA2, PAES, SMPSO and CellDE increase but not signicantly when the number of features increases.
For NSGA-II, MOCell and RS, the performances decrease but not signicantly as the increasing number of features since the Spearmans
is less than 0 and p-values are greater than 0.05. In summary, we

conclude that RWGA has the ability to solve a wide range of problems with gradually better performance and the other seven algorithms are not inuenced signicantly by the number of features
(Table 16).
Increasing Number of Associated Test Cases: Based on the results, we
observed that for WBGA_W1 , WBGA_W2 , WBGA-MO, RWGA, SPEA2,
PAES and SMPSO, the values of MFV_TC increase but not signicantly
as the number of associated test cases increases since Spearmans
is greater than 0 and p-values are greater than 0.05, i.e., the performances of WBGA_W1 , WBGA_W2 , WBGA-MO, RWGA, SPEA2, PAES
and SMPSO increase but not signicantly with the growth of associated test cases. For the other algorithms (NSGA-II, MOCell, CellDE
and RS), the values of MFV_TC decrease but not signicantly when
the associated test cases increase, showing the performances of these
algorithms decrease but not signicantly as the increasing number of
test cases. In summary, we can conclude even with the increase in
the number of test cases, the performance of search algorithms is not
inuenced as shown in Table 16.
ARTICLE IN PRESS
JID: JSS
16
[m5G;December 8, 2014;18:52]
Table 14
Results for comparing the eight algorithms for OFV.
Pair of algorithms
RWGA vs. NSGA-II

RWGA vs. MOCell
RWGA vs. SPEA2
RWGA vs. PAES
RWGA vs. SMPSO
RWGA vs. CellDE
RWGA vs. RS
NSGA-II vs. MoCell
NSGA-II vs. SPEA2
NSGA-II vs. PAES
NSGA-II vs. SMPSO
NSGA-II vs. CellDE
NSGA-II vs. RS
MOCell vs. SPEA2
MOCell vs. PAES
MOCell vs. SMPSO
MOCell vs. CellDE
MOCell vs. RS
SPEA2 vs. PAES
SPEA2 vs. SMPSO
SPEA2 vs. CellDE
SPEA2 vs. RS
PAES vs. SMPSO
PAES vs. CellDE
PAES vs. RS
SMPSO vs. CellDE
SMPSO vs. RS
CellDE vs. RS
C20
C40
C60
C90
0.51
0.57
0.55
0.61
0.71
0.56
0.66
0.59
0.61
0.58
0.45
0.56
0.64
0.65
0.58
0.59
0.49
0.71
0.34
0.28
0.45
0.55
0.48
0.46
0.65
0.49
0.63
0.72
0.22
0.03
0.07
0.02
0.01
0.14
0.01
0.11
0.04
0.15
0.33
0.25
0.02
0.02
0.08
0.14
0.21
0.02
0.02
0.01
0.11
0.32
0.17
0.21
0.02
0.32
0.01
0.01
0.62
0.57
0.62
0.57
0.67
0.75
0.79
0.51
0.69
0.56
0.61
0.57
0.69
0.73
0.55
0.56
0.59
0.75
0.31
0.32
0.32
0.51
0.51
0.45
0.72
0.46
0.77
0.75
0.01
0.02
0.01
0.11
0.03
0.01
0.02
0.23
0.04
0.17
0.04
0.08
0.01
0.04
0.12
0.13
0.17
0.01
0.02
0.02
0.03
0.24
0.09
0.14
0.02
0.29
0.01
0.01
0.59
0.72
0.69
0.71
0.64
0.60
0.65
0.61
0.65
0.61
0.52
0.51
0.58
0.55
0.51
0.52
0.60
0.67
0.43
0.38
0.28
0.48
0.43
0.52
0.66
0.43
0.64
0.68
0.04
0.01
0.03
0.01
0.02
0.06
0.02
0.09
0.09
0.07
0.32
0.29
0.01
0.09
0.21
0.08
0.19
0.04
0.09
0.02
0.02
0.12
0.17
0.15
0.03
0.16
0.03
0.02
0.52
0.54
0.64
0.56
0.55
0.68
0.72
0.53
0.78
0.45
0.55
0.63
0.70
0.64
0.57
0.51
0.53
0.66
0.29
0.41
0.27
0.54
0.52
0.41
0.71
0.38
0.69
0.75
0.11
0.12
0.04
0.08
0.09
0.04
0.02
0.19
0.02
0.22
0.18
0.04
0.02
0.04
0.09
0.14
0.19
0.02
0.01
0.07
0.02
0.08
0.25
0.09
0.01
0.06
0.02
0.01
6.2.3. Timing analysis for RWGA and the other selected search
algorithms
Similarly, we also performed a timing analysis for the best weightbased GAs (i.e., RWGA) and the other kinds of selected search algorithms (e.g., NSGA-II) as compared with RS. Table 17 shows the
Table 15
Results for the Vargha and Delaney statistics-articial problems (without/with Whitney U test).
RQ
Pair of Algorithms (A vs. B)
A>B
A<B
A=B
RQ2
RWGA vs. RS
NSGA-II vs. RS
MOCell vs. RS
SPEA2 vs. RS
PAES vs. RS
SMPSO vs. RS
CellDE vs. RS
489/480
467/449
438/419
298/251
430/414
388/362
423/399
11/0
33/14
62/46
202/179
70/59
112/93
77/65
0/20
0/37
0/35
0/70
0/27
0/45
0/36
RQ3
RWGA vs. NSGA-II

RWGA vs. MOCell
RWGA vs. SPEA2
RWGA vs. PAES
RWGA vs. SMPSO
RWGA vs. CellDE
NSGA-II vs. MOCell
NSGA-II vs. SPEA2
NSGA-II vs. PAES
NSGA-II vs. SMPSO
NSGA-II vs. CellDE
MOCell vs. SPEA2
MOCell vs. PAES
MOCell vs. SMPSO
MOCell vs. CellDE
SPEA2 vs. PAES
SPEA2 vs. SMPSO
SPEA2 vs. CellDE
PAES vs. SMPSO
PAES vs. CellDE
SMPSO vs. CellDE
432/405
403/397
464/451
381/362
453/438
380/357
269/252
398/377
266/242
226/219
245/227
395/374
302/281
267/243
201/182
170/145
141/107
179/152
223/201
198/182
244/219
68/57
97/85
36/23
119/104
47/39
120/103
231/204
102/86
234/219
273/241
255/239
105/91
198/167
233/200
299/283
330/307
359/316
321/303
277/255
302/270
256/238
0/32
0/18
0/26
0/24
0/23
0/40
0/44
0/47
0/39
0/40
0/34
0/35
0/52
0/57
0/35
0/48
0/77
0/45
0/46
0/48
0/43
average running time per run taken by each selected search algorithms. We also obtained a p-value of 0.63 with the Kruskal
Wallis test suggesting there are no signicant differences between
the running time of the selected search algorithms as compared
with RS. Based on the results, we can conclude that all the selected search algorithms did not take signicantly different time
for running as compared with RS in terms of nding an optimal
solution.
7. Discussion
In this section, we discuss the results of the experiments based
on our industrial case study and the 500 articial problems. Section 7.1 provides insight on the results of RQ1RQ3, while Section
7.2 discusses results of RQ4. Recall that RQ1RQ3 address questions
related to cost/effectiveness and performance of our minimization
problem (as compared with RS), while RQ4 refers to the impact of the
increasing number of features and the number of test cases on the
performance of the selected multi-objective search algorithms (Section 5.1.1). Before getting deeper into the specic discussion related
with the research questions, the key results are rst summarized as
below:
Different algorithms achieve the best performance for each objective, i.e., CellDE for TMP, MOCell for FPC, NSGA-II for FDC and
SPEA2 for OET;
RWGA achieves signicantly best performance among all the selected search algorithms when considering all the objectives together;
The performance of weight-based GAs (e.g., RWGA) increases with
the growth of features, but is not signicantly inuenced by the
increasing number of associated test cases;
All the other search algorithms are not signicantly impacted by
the increasing complexity of problems (increasing features and
associated test cases).
ARTICLE IN PRESS
JID: JSS
[m5G;December 8, 2014;18:52]
17
Table 16
Spearmans correlation analysis for the select algorithms with the increasing number of features and associated test cases.
Algorithms
Increasing number of features
WBGA_W1
WBGA_W2
WBGA-MO
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
Increasing number of associated test cases
Spearman
Prob > | |
Spearman
Prob > | |
0.52
0.47
0.69
0.74
0.13
0.11
0.09
0.02
0.15
0.08
0.08
<0.0001
<0.0001
<0.0001
<0.0001
0.5546
0.6852
0.8019
0.7980
0.5742
0.7006
0.8704
0.08
0.10
0.15
0.19
0.10
0.02
0.12
0.14
0.07
0.13
0.09
0.5764
0.5123
0.4892
0.4719
0.6714
0.7653
0.5417
0.4928
0.6986
0.3985
0.7619
7.1. Discussion related to RQ1RQ3

Weight-based GAs require sets of weights (e.g., w1 , w2 and w3 )
that can be provided by domain experts based on the testing requirements. Our results and analysis show that RWGA presents the
best performance among the three weight-based GAs and RS. In fact,
RWGA assigns weights dynamically during the search and thus the
search is more eciently guided toward the best weights, to reach
the threshold values for TMP, FPC, FDC and OET.
When comparing the best weight-based GA (RWGA) and the other
search algorithms, from our industrial case study, we observe that the
solutions obtained by RWGA, NSGA-II, MOCell, PAES, SMPSO, CellDE
can reach the expected thresholds for each objective (Section 6.2.1).
SPEA2 managed to reach the expected thresholds for TMP and OET
but not for FPC and FDC. RS cannot nd optimal solutions for the
thresholds (RQ1). When comparing the selected search algorithms
with RS, all the algorithms except SPEA2 have signicantly better
performance than RS in nding optimal solutions for each objective
function and OFV considering all the objectives. SPEA2 has signicantly better performance than RS for TMP and OET but has almost
equivalent performance as RS for FPC, FDC and OFV (RQ2).
Moreover, when studying the performance of each individual algorithm, different algorithms achieve the best performance for different
objectives (e.g., CellDE for TMP, MOCell for FPC, NSGA-II for FDC and
SPEA2 for OET). However, when combining all objectives together,
RWGA has signicantly better performance than all the other search
algorithms (RQ3). For the articial problems (RQ1RQ3), we observe
consistent results with our observations on the industrial case study.
This difference in performance can be explained as follows: RWGA
aims at balancing all the required objectives and nding an optimal solution by considering all the objectives together. Moreover, by
dynamically changing weights at each generation, RWGA more thoroughly explores the search space of weights. This dynamic assignment
of the weights (i.e., RWGA) is promising to nd an optimal solution in
a given number of generations, as compared to a strategy using only
xed weights for all the generations (e.g., WBGA_W1 and WBGA_W2 ).
Recall from Section 2 that NSGA-II, MOCell, SPEA2, PAES, SMPSO and
CellDE are used to seek a set of optimal non-dominated solutions.
However, when considering all the objectives together (OFV), such
non-dominated solutions may be worse than the solutions found
by RWGA that aims at balancing all the objectives together. Notice
that the solution obtained by RWGA may not dominate the solutions
obtained by the other search algorithms (e.g., NSGA-II) but RWGA

can manage to nd the best solution with maximum overall tness
function (OFV) for the minimization problem. However, if users only
concern one of the most important objectives (e.g., FDC), our approach
with corresponding tool (Section 8) will recommend the most suitable
algorithm accordingly (e.g., MOCell for FDC).
7.2. Discussion related to RQ4
When looking into the performance of each individual algorithm
(RQ4) as the number of features and test cases increase (MFV_F
and MFV_TC as shown in Figs. 2 and 3), we observe that RWGA
presents the best performance and has the ability to solve a wider
range of problems with gradually better performance (Fig. 2). In
Fig. 2. Mean tness values for number of features.
Table 17
Average running time for the selected weight-based search algorithms.
Algorithms
RWGA
NSGA-II
MOCell
SPEA2
PAES
SMPSO
CellDE
RS
Time (s)
1.37
1.54
1.23
1.98
2.06
2.19
1.45
1.13
Fig. 3. Mean tness values for number of test cases.
JID: JSS
18
ARTICLE IN PRESS
[m5G;December 8, 2014;18:52]
addition, when the number of test cases associated to each feature

increases, the performance of RWGA is not signicantly impacted
(Fig. 3).
To show the combined effect of the increase in the number of
features and test cases on the performance of search algorithms,
we show surface plots based on the mean tness value (MFV,
Section 5.1.3) in Fig. 4. From Fig. 4, we observe that RWGA reaches
the best performance among all the algorithms. Moreover, the performance of WBGA and WBGA-MO also increases as the number of
features increases. Finally, the performance of the other seven algorithms is not signicantly impacted (notice that this observation is
consistent with the results presented in Figs. 2 and 3). For instance,
the performance of NSGA-II stays stable as the increasing number of
the features and associated test cases. It is interesting to observe that
the performance of SPEA2 remains worse than the other search algorithms without signicantly changing with the growth of features
and associated test cases. Notice that RS performs worse among all the
selected search algorithms. These observations suggest that the selected search algorithms together with our proposed tness function
are scalable and sucient for test minimization for the products of
varying complexity. Sections 7.2.1 and 7.2.2 further provide detailed
explanations about the performance when the number of features
and test cases increases.
7.2.1. Increasing number of features
Three different types of solution spaces need to be rst introduced
to understand the rest of the section:
(1) The Solution Space (black ellipse in Fig. 5) refers to the space of
potential solutions for any search algorithm, depending on the
number of test cases. When the number of features increases,
then the solution space can either: (A) increase when more
test cases associated with new features are introduced; or (B)
remain the same when no more test cases are introduced to the
solution space (i.e., the existing test cases can cover the new
features). For instance, in our industrial case study (Cisco), two
call protocols H323 and SIP are designed and relevant test cases
are developed for testing H323 and SIP. If a new call protocol
based on H323 and SIP is designed, the existing test cases may
be reused for testing such new protocol. In this case, no new
test cases will be introduced to the solution space.
(2) The Balanced Solution Space (red ellipse in Fig. 5) represents
the space for optimal solutions, balancing all the objectives
together. Weight-based GAs aims at guiding the search within
this subspace in order to nd optimal balanced solutions. When
the number of features increases, the balanced solution space
can either (A) grow, (B) remain the same or (C) decrease.
(3) The Non-dominated Solution Space (green ellipse in Fig. 5) refers
to the space for non-dominated optimal solutions. The algorithms based on dominance theory (i.e., NSGA-II, MOCell,
SPEA2, PAES, SMPSO and CellDE) can guide the search to nd a
set of non-dominated solutions within this solution space. Similarly, the non-dominated solution space can either (A) grow,
(B) remain the same, or (C) decrease with the growth of features.
Based on the above-mentioned denitions, our observations for
experiments can be explained under the situation when the solution space, the balanced solution space and the non-dominated solution all grow as the increasing number of features. More specically, the gradual improvement of performance of weight-based GAs
(WBGA, WBGA-MO and RWGA) is interpreted as follows: Weightbased GAs (e.g., RWGA) guides the search to nd optimal balanced solutions, considering all the objectives. Hence, the space of
balanced solutions grows when the space for potential solutions
grows. As a result, weight-based GAs (i.e., RWGA) easily nds an
Fig. 4. Surface plots for all the selected algorithms.
ARTICLE IN PRESS
JID: JSS
Balanced
Non-dominated
[m5G;December 8, 2014;18:52]
19
Concluding remarks: When multi-objective test minimization for

software product lines is sought, we recommend using CellDE, MOCell, NSGA-II and SPEA2 for the objectives TMP, FPC, FDC and OET,
respectively, when one objective has more importance than the others. In contrast, using RWGA is recommended when optimizing all
the objectives together is sought. The latter situation happened in
our industrial case study.
Fig. 5. An overall diagram for three types of space.
8. Automation
optimal balanced solution from a larger balanced solution space with
the aim for balancing all the objectives together.
For the other algorithms based on the dominance theory (i.e.,
NSGA-II, MOCell, SPEA2, PAES, SMPSO and CellDE), the performance
is not signicantly inuenced by increasing number of features since
these algorithms guide the search in nding a set of non-dominated
solutions (usually one objective is the main concern). The overall tness value (OFV, Section 5.1.3) does not change signicantly even
when the space for non-dominated solutions increases, as shown in
(a) and (b) of Fig. 5. For instance, solution a and solution b are two
obtained non-dominated solutions by NSGA-II with TMP = 0.8, FPC =
0.5, FDC = 0.5, OET = 0.5, AEF = 0.5 and TMP = 0.5, FPC = 0.8, FDC = 0.5,
OET =0.5, AEF = 0.5, respectively, i.e., the main objective for solution a
is the percentage of test minimization (TMP) while the main concern
for solution b is feature pairwise coverage (FPC). But for the overall
tness value (OFV) when considering all the objectives, solution a and
b are equivalent since the OFV values for both solutions are the same
(i.e., both values of OFV for solution a and solution b are 0.56, Section
5.1.3), which is the main reason the performance of NSGA-II, MOCell, SPEA2, PAES, SMPSO and CellDE stays stable as shown in Figs. 2
and 4.
For random search (RS), the performance remains the worst and
is not inuenced by the increasing features since it chooses solutions
randomly without any guideline.
For the other combinations for the solution space, the balanced
solution space and the non-dominated solution, the observations for
experiments may not be the same as what we observed in our experiments. This further requires more focused experiments that assess
the performance of the search algorithms with varying combinations
of increase, decrease, and no change in the three types of solution
spaces.
7.2.2. Increasing number of associated test cases

When increasing the number of test cases, we observed no signicant impact on the performance for all the selected algorithms, which
can be explained as follows: The complexity of problems mainly depends on the number of features (FPC measures how many pairs are
covered, but the number of test cases is ignored) and the relationships between features and test cases are not complicated in our case
(i.e., the only relationship between features and associated test cases
is that one feature can be associated with a certain number of test
cases as discussed in Section 4). As a result, when the complexity of a
problem is mainly determined by the number of features, the number
of associated test cases may have no signicant impact on the performance of algorithms as shown in Figs. 3 and 4. However, in other cases,
more complicated relationships may exist between features and associated test cases, e.g., testing some features requires several steps
and each step is covered by a certain number of test cases or execution
of test cases depends on the execution of other test cases. Thus, the
number of associated test cases may have signicant impact on the
performance of algorithms. We plan to further investigate this problem by using different case studies and other focused experiments in
the future.
In this section, we present the tool support for test minimization using search algorithms along with our proposed tness functions. The tool is called TEst Minimization with Search Algorithms
(TEMSA) developed to minimize the test cases at the same time
improving effectiveness and reducing predetermined cost based
on various cost/effectiveness measures. TEMSA is implemented as
a web-based application on top of the jMetal library using Java Sencha ExtJS (a JavaScript Framework for Rich Web Apps) (Sencha, 2010).
The architecture of TEMSA is shown in Fig. 6 (TEMSA can be tried out
at the website: http://zen-tools.com/TEMSA/).
The input of TEMSA is a set of selected features and test cases for
testing a product as an XML le in our industrial case study. To use
TEMSA, users have to provide an input le corresponding to our specic XML schema (the related XML schema and sample input XML le
can be downloaded from our tool website). Cost measures (e.g., OET)
and effectiveness measures (e.g., TMP, FPC, FDC and AEF) are integrated
into TEMSA for covering various objectives. Other objective measures
can also be coded into TEMSA by formulating them as specic objective functions mathematically as we discussed in Section 3.2. For the
test cases, each one includes three key attributes for cost/effectiveness
measures, i.e., SucR, EF and AET (Section 3.2), which are obtained from
test execution history. Based on these inputs, TEMSA outputs a set of
minimized test cases organized as another XML le which also conforms to our specic XML schema and can be transformed into any
other schema when needed. In our context, the output can be put into
test execution engine for scheduling and further executing for System
Under Test (SUT). Afterwards, the output of executed test cases will
be an input to update test execution history in the repository, such as
SucR, EF and AET for the executed test cases.
Moreover, our tool recommends different algorithms depending on test minimization objectives based on the results of our
empirical study as shown in Fig. 7. First, a user selects an objective
with highest importance (Activity A1) and TEMSA will automatically
recommend the relevant best algorithm based on the selected objective. For instance, if the highest importance is to reduce the number
of test cases (TMP), then CellDE will be recommended by TEMSA (Activity A2). Similarly, if feature pairwise coverage (FPC), fault detection
capability (FDC) and the allowed time budget (OET) are the most important objectives, TEMSA will recommend users to adapt MOCell,
NSGA-II and SPEA2 for their test minimization, respectively (Activity
A3, A4 and A5). Moreover, if a user wants to balance all the objectives together for test minimization, they do not need to select any
objective, (Activity A1) and RWGA will be recommended by TEMSA
automatically (Activity A6).
9. Threats to validity
In this section, we discuss four threats to validity for our experiments and how we address them.
9.1. Internal validity

Internal validity threats exist when the outcome of results
is inuenced by internal factors (e.g., parameters of search
JID: JSS
20
ARTICLE IN PRESS
[m5G;December 8, 2014;18:52]
Fig. 6. Architecture for TEMSA
algorithms) and are not necessarily due to the application of the treatment (e.g., search algorithms) being studied (Ali et al., 2010; Barros
and Neto, 2011). A possible threat to internal validity is that we have
experimented with only one-default conguration settings for the
parameters of the selected search algorithms. However, these settings are in accordance with the common guidelines in the literature
and our previous experience on testing problems (Arcuri and Briand,
2011; Sheskin, 2007; Wang et al., 2013c).
cost that have severe limitations, as they are not precise. To reduce
construct validity threats in our context, we used the same stopping
criteria for all algorithms, i.e., number of tness evaluations. We ran
each algorithm for 25,000 evaluations to seek the best solution for
test minimization. This criterion is a comparable measure across all
the algorithms since each iteration requires updating the obtained
solution and comparing the computed value of tness function.
9.4. External validity
9.2. Conclusion validity

Conclusion validity threats are concerned with factors that can
inuence the conclusion that can be drawn from the results of
the experiments. The most probable conclusion validity threat in
experiments involving randomized algorithms is due to random
variations. To address it, we repeated experiments 100 times to reduce the possibility the results were obtained by chance. Furthermore, to determine the probability of yielding higher performance
by different algorithms, we measured the effect size using Vargha
and Delaney statistics test since it is appropriate for non-parametric
effect size measure, which matches our situation. Meanwhile, we
performed MannWhitney U test to determine the statistical significance of the results (Arcuri and Briand, 2011). Finally, Spearmans
rank correlation coecient is used to measure the potential impact on the performance of selected search algorithms by increasing number of features and associated test cases since it is mainly
used for non-parametric correlation coecient measure and suits our
objective in this paper (Yue et al., 2013).
9.3. Construct validity
For the construct validity threats, we looked at the validity of
the comparison measures for all the selected search algorithms.
The most frequently observed threat was using some measures of
External validity threats exist when the outcome of results is inuenced by external factors and are not necessarily due to the application of the treatment being studied (Ali et al., 2010; Barros
and Neto, 2011). One common external validity threat in the software engineering experiments is about generalization of results. One
may argue the results cannot be generalized for other case studies if we ran our experiments on just one industrial case study.
However, such threat to external validity is common to all empirical studies. In addition, we also conducted an empirical evaluation
using carefully designed 500 articial problems and the results are
consistent with the industrial case study.
10. Related works

In the context of regression testing, the optimization problem
can be classied into the following three problems, i.e., test case
selection, minimization, and prioritization (Fischer et al., 1981;
Harman, 2011; Yoo and Harman, 2007, 2012; Zheng et al., 2007).
More specically, test case selection focuses on selecting a subset
of test cases from the existing test suite with aim to test a modied version of systems or programs while test minimization eliminates the redundant test cases from the existing test suite for the
current systems or programs in order to reduce the cost of testing
Fig. 7. Recommendations for algorithms by TEMSA
JID: JSS
ARTICLE IN PRESS
(e.g., time). Test prioritization orders the test cases in the existing
test suite to test the current systems or programs with the objective of achieving the pre-dened criteria (e.g., maximum number of executing test cases) in a given time budget. The main difference between selection and minimization is that test minimization eliminates the redundant test cases permanently for the systems (programs), while test case selection refers to select relevant
test cases temporarily for testing the modied version of the systems (Harman, 2011; Yoo and Harman, 2012). However, from the
perspective of optimization, there is no signicant difference between test case selection and test minimization (Harman, 2011), i.e.,
both of them aim at choosing a subset of test cases from the existing test suite at the same time achieving pre-dened objectives
for optimization, which was called test case selection problem in
(Harman, 2011).
Our work is related to the above-mentioned selection problem,
but as opposed to the existing works, our work lies in the context of product line testing (e.g., feature pairwise coverage is used
to measure the effectiveness of the minimized test suite based on
the coverage of feature pairs from the product line view rather
than code level (e.g., code coverage) in the most of the existing
works for regression testing). Although there exist a large number
of test selection/minimization techniques in the context of regression testing, there is not enough evidence that these techniques can
be adapted to product lines (e.g., whether code coverage in regression testing is equivalent to the feature coverage in product line
testing).
Yoo and Harman (2007) used a greedy algorithm and a multiobjective search algorithm NSGA-II for test case selection for the following three measures: code coverage, fault detection history, and
execution time in the context of regression testing. Notice that these
three measures are similar to our cost/effectiveness measures FPC,
FDC and OET. The difference with our work is that rst we select a
set of relevant test cases for a new product using feature models as
we discussed in (Wang et al., 2012, 2013a) and then we perform test
minimization based on two additional measures that are TMP and
AEF. In addition, we conducted an empirical study to compare ten
multi-objective algorithms belonging to four different mechanisms
and evaluate their performance, which was not addressed in Yoo
and Harman (2007).
In another work related to test prioritization in regression testing (Walcott et al., 2006), a two-objective problem (i.e., code coverage and execution time) is converted into a single-objective problem using an arithmetical combination of weights for the tness
function. These two objectives are close to our cost/effectiveness
measures FPC and OET. However, as compared with code coverage, our dened FPC mainly focuses on the coverage of feature pairs from the view of product line rather than code level
of the systems. Moreover, additional three measures are dened
mathematically in our work and we do not limit our scope
within weight-based GAs, i.e., different types of search algorithm
are empirically evaluated, e.g., NSGA-II and SPEA2.
As compared with our previous work (Wang et al., 2013c),
we dene two additional cost/effectiveness measures (i.e., AEF and
OET) for test minimization problem. Moreover, we further compare the performances for the best weight-based GA (i.e., RWGA)
and the other six different search algorithms using an industrial case study and 500 designed articial problems, which cover
various kinds of search algorithms and make the results and
analysis more persuasive.
To the best of our knowledge and existing published reviews
(Harman et al., 2009; Yoo and Harman, 2012), none of the existing works studied test minimization in the context of product line
considering TMP, FPC, FDC, OET and AEF all together. In addition, it is
one of the few works, which extensively compares the performance
of search algorithms belonging to different mechanisms (e.g., evolu-
[m5G;December 8, 2014;18:52]
21
tionary algorithms, swarm algorithms, and hybrid algorithms) for test

suite minimization in the context of product lines.
11. Conclusion and future work
In this paper, we have proposed and have evaluated a tness function in conjunction with ten multi-objective search algorithms, to
minimize test suites in the context of product line testing. Our tness function takes effectiveness measures (i.e., Test Minimization
Percentage (TMP), Feature Pairwise Coverage (FPC), Fault Detection
Capability (FDC) and Average Execution Frequency (AEF)) and cost
measures (i.e., Overall Execution Time (OET)) into account to guide
the search. The search algorithms, along with our tness function,
are evaluated on an industrial case study and 500 articial problems of varying size and complexity using a tool we have developed
for this purpose, called TEst Minimization with Search Algorithms
(TEMSA).
Given a set of thresholds for each objective, the results show that,
unlike RS, WBGA-MO, RWGA, NSGA-II, MOCell, PAES, SMPSO and
CellDE can reach these thresholds, while SPEA2 achieves acceptable
performance for TMP/OET, which is not the case for FPC/FDC. Meanwhile, we observe that the performance of WBGA (i.e., WGBA_W1
and WGBA_W2 ) is not stable since it largely depends on the predetermined weights for the objectives. When comparing the selected algorithms, our results show that WBGA, WBGA-MO, RWGA,
NSGA-II, MOCell, PAES, SMPSO and CellDE signicantly outperform
RS and SPEA2 has almost equivalent performance than RS. From
the results on articial problems, we conclude that RWGA also
achieves the best performance when taking all objectives into account. Moreover, RWGA can solve a wider range of problems and
its performance is improved when the number of features number is growing. Notice the performance of other weight-based GA
is also signicantly increased with the growth of features. For rest
of the algorithms, the performance is not signicantly inuenced by
the increasing number of features or test cases, when it comes to
nding an optimal solution.
Our short-term future plan includes the analysis of another industrial case study in order to complement our experimental results and rene the design of TEMSA. We also plan to rene the
tness function by considering other objectives, such as the number
of available computing resources, and evaluate other major search
algorithms, such as Archive-Based hYbrid Scatter Search (AbYSS),
which combines both scatter search and genetic algorithms.
Acknowledgements
We gratefully thank Marius Liaaen and his team at Cisco for providing us the valuable industrial case study. All the authors acknowledge
funding from the Research Council of Norway.
References
Ali, S., Briand, L.C., Hemmati, H., Panesar-Walawege, R.K., 2010. A systematic review of
the application and empirical investigation of search-based test case generation.
IEEE Trans. Softw. Eng. 36, 742762.
Arcuri, A., Briand, L., 2011. A practical guide for using statistical tests to assess randomized algorithms in software engineering, In: Proceedings of the 33rd International Conference on Software Engineering. ACM, Waikiki, Honolulu, HI,
USA, pp. 110.
Barros, M.d.O., Neto, A.C.D., 2011. Threats to Validity in Search-based Software Engineering Empirical Studies. UNIRIO Universidade Federal do Estado do Rio de
Janeiro 0006/2011.
Benavides, D., Segura, S., Ruiz-Corts, A., 2010. Automated analysis of feature models
20 years later: a literature review. Inf. Syst. 35, 615636.
Brownlee, J., 2012. Clever Algorithms: Nature-Inspired Programming Recipes. lulu.com,
1st ed..
CiscoSystems, C., 2010. Cisco Telepresence Codec c90, Data Sheet..
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective
genetic algorithm: NSGA-II. IEEE Trans. Evol. Comput. 6, 182197.
JID: JSS
22
ARTICLE IN PRESS
[m5G;December 8, 2014;18:52]
Durillo, J., Nebro, A., Luna, F., Alba, E., 2008. Solving Three-objective optimization problems using a new hybrid cellular genetic algorithm. In: Rudolph, G., Jansen, T.,
Lucas, S., Poloni, C., Beume, N. (Eds.), Parallel Problem Solving from Nature PPSN
X. Springer, Berlin Heidelberg, pp. 661670.
Durillo, J.J., Nebro, A.J., 2011. jMetal: a Java framework for multi-objective optimization.
Adv. Eng. Softw. 42, 760771.
Engstrm, E., Runeson, P., 2011. Software product line testing a systematic mapping
study. Inf. Softw. Technol. 53, 213.
Fischer, K.F., Raji, F., Chruscicki, A., 1981. A methodology for retesting modied software, National Telecommunications Conference (NTC), pp. 16.
GmbH, P., 2006. Variant Management with Pure::Variants. Technical White Paper.
Greer, D., Ruhe, G., 2004. Software release planning: an evolutionary and iterative
approach. Inf. Softw. Technol. 46, 243253.
Harman, M., 2011. Making the case for MORTO: multi objective regression test optimization. In: Proceedings of the 2011 IEEE Fourth International Conference on
Software Testing, Verication and Validation Workshops. IEEE Computer Society,
pp. 111114.
Harman, M., Mansouri, S.A., Zhang, Y., 2009. Search Based Software Engineering: A Comprehensive Analysis and Review of Trends Techniques and Applications. Technical
Report TR-09-03.
Knowles, J.D., Corne, D.W., 2000. Approximating the nondominated front using the
pareto archived evolution strategy. Evol. Comput. 8, 149172.
Konak, A., Coit, D.W., Smith, A.E., 2006. Multi-objective optimization using genetic
algorithms: a tutorial. Reliab. Eng. Syst. Saf. 91, 9921007.
Kuo-Chung, T., Yu, L., 2002. A test generation strategy for pairwise testing. IEEE Trans.
Softw. Eng. 28, 109111.
Lopez-Herrejon, R.E., Chicano, F., Ferrer, J., Egyed, A., Alba, E., 2013. Multi-objective optimal test suite computation for software product line pairwise testing, International
Conference on Software Maintenance (ICSM), pp. 404407.
Mcdonald, J.H., 2009. Handbook of Biological Statistics.
McGregor, J., 2001. Testing a Software Product Line (CMU/SEI-2001-TR-022). Software
Engineering Institute, Carnegie Mellon University, Pittsburgh, PA.
Nebro, A., Durillo, J., Luna, F., Dorronsoro, B., Alba, E., 2007. Design issues in a multiobjective cellular genetic algorithm. In: Obayashi, S., Deb, K., Poloni, C., Hiroyasu, T.,
Murata, T. (Eds.), Evolutionary Multi-Criterion Optimization. Springer, Berlin Heidelberg, pp. 126140.
Nebro, A.J., Durillo, J.J., Garcia-Nieto, J., Coello Coello, C.A., Luna, F., Alba, E., 2009. SMPSO:
a new PSO-based metaheuristic for multi-objective optimization, MCDM09. IEEE
Symposium on Computational intelligence in Multi-criteria Decision-Making,
2009, pp. 6673.
Runeson, P., Engstrm, E., 2012. Regression testing in software product line engineering.
Adv. Comput. 86, 223263.
Sencha, 2010. http://www.sencha.com/products/extjs..
Sheskin, D.J., 2007. Handbook of Parametric and Nonparametric Statistical Procedures.
Chapman&Hall&CRC.
Walcott, K.R., Soffa, M.L., Kapfhammer, G.M., Roos, R.S., 2006. TimeAware test suite
prioritization, In: Proceedings of the 2006 International Symposium on Software
Testing and Analysis. ACM, Portland, Maine, USA, pp. 112.
Wang, S., Ali, S., Gotlieb, A., 2013. Minimizing test suites in software product lines using
weight-based genetic algorithms, In: Proceeding of the Fifteenth Annual Conference on Genetic and Evolutionary Computation Conference. ACM, Amsterdam, The
Netherlands, pp. 14931500.
Wang, S., Gotlieb, A., Ali, S., Liaaen, M., 2013. Automated test case selection using
feature model: an industrial case study. In: Moreira, A., Schtz, B., Gray, J., Vallecillo, A., Clarke, P. (Eds.), Model-Driven Engineering Languages and Systems.
Springer, Berlin Heidelberg, pp. 237253.
Wang, S., Gotlieb, A., Liaaen, M., Briand, L.C., 2012. Automatic selection of test execution
plans from a video conferencing system product line, In: Proceedings of the VARiability for You Workshop: Variability Modeling Made Useful for Everyone. ACM,
Innsbruck, Austria, pp. 3237.
Yoo, S., Harman, M., 2007. Pareto ecient multi-objective test case selection, In: Proceedings of the 2007 International Symposium on Software Testing and Analysis.
ACM, London, United Kingdom, pp. 140150.
Yoo, S., Harman, M., 2012. Regression testing minimization, selection and prioritization:
a survey. Softw. Test. Verif. Reliab. 22, 67120.
Yue, T., Briand, L.C., Labiche, Y., 2013. Facilitating the transition from use case models
to analysis models: approach and experiments. ACM Trans. Softw. Eng. Methodol.
22, 138.
Zhang, Y., Finkelstein, A., Harman, M., 2008. Search based requirements optimisation: existing work and challenges. In: Paech, B., Rolland, C. (Eds.), Requirements
Engineering: Foundation for Software Quality. Springer, Berlin Heidelberg, pp. 88
94.
Zheng, L., Harman, M., Hierons, R.M., 2007. Search algorithms for regression test case
prioritization. IEEE Trans. Softw. Eng. 33, 225237.
Zitzler, E., Laumanns, M., Thiele, L., 2001. SPEA2: improving the strength
pareto evolutionary algorithm, The EUROGEN 2001-Evolutionary Methods for
Design, Optimization and Control with Applications to Industrial Problems, pp.
95100.
Shuai Wang obtained his master degree for computer science and technology in the eld of software testing from
Beihang University, Beijing, China. He is currently working as PhD student at the Certus Software Verication &
Validation Center hosted by Simula Research Laboratory
and the Department of Informatics, University of Oslo, Norway. His main research interests are: variability modeling,
model-based testing and search-based testing. He is a student member of the ACM and IEEE computer society.
Shaukat Ali is currently a research scientist in Certus Software Verication & Validation Center, Simula Research
Laboratory, Norway. He has been aliated to Simula Research Lab since 2007. He has been involved in many industrial and research projects related to Model-Based Testing (MBT) and Empirical Software Engineering since 2003.
He has experience of working in several industries and
academia research groups in many countries including UK,
Canada, Norway and Pakistan. He is a member of the IEEE
Computer Society.
Arnaud Gotlieb obtained his PhD degree in the elds of Software Testing and Constraint Programming from the University of Nice-Sophia Antipolis, in 2000. He has worked
7 years in the Industry at THALES (pre. DASSAULT Electronics) where he was successively PhD student, engineer
and project manager in the eld of Software Testing. From
2002 to 2011, he has been worked in Rennes as an INRIAs researcher on Constraint-Based Testing. Since 2011,
he joined the Simula Research Laboratory as a senior research scientist and from 2012, he leads the CERTUS Software Validation & Verication Center and heads the Software Engineering department of SIMULA. He is the main
author of more than forty publications and coauthor of
more than seventy, and he is the main architect of several constraint solving engines targeted to the testing of critical programs. He is a member of the IEEE
Computer Society.

Wang 2014

Uploaded by

Copyright:

Available Formats

Wang 2014

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wang 2014

Uploaded by

Copyright:

Available Formats

ARTICLE IN PRESS

The Journal of Systems and Software 000 (2014) 122

Contents lists available at ScienceDirect

The Journal of Systems and Software

Cost-effective test suite minimization in product lines using search

An overall view of our product line testing project is presented

Fig. 1. An overview of cost-effective testing of products in a product line.

(Brownlee, 2012), which are: Evolutionary Algorithms (EAs), Swarm

2.1. Overview of search algorithms

2.2. Evolutionary Algorithms (EAs)

value for each solution is calculated by summing up a strength raw

formal representation of the search problem, followed by denitions

2.3. Swarm Algorithm

product pi which is to be tested (Benavides et al., 2010), where fi

3.1. Problem representation

3.1.1. Basic concepts

Particle Swarm Optimization (PSO) is a biological metaheuristic

3.1.2. Cost/effectiveness measures

3.1.3. Problem representation

Eff ect (sk , eff ectj )

Eff ect (sl , eff ectj )

Cost (sk , cos tj )

Cost (sl , cos tj )

3.2. Denitions and functions for effectiveness/cost measures

As discussed in Section 3.1, ntsk is the number of test cases for

Num_FPsk is the number of feature pairs covered in the test cases

sk refers to a solution for a product including a set of test cases and

ntsk is the number of test cases for the solution sk , where 1

ntsk is the number of test cases for the solution sk , where 1

EFtci is the execution frequency of the test case tci in a given

Fitness = f (TMP, FPC, FDC, OET, AEF )

Fitness_w = 1 w1 nor(TMP ) w2 nor(FPC ) w3

at Cisco, and test execution history. Meanwhile, there is no specic

(nor(TMP) + nor(FPC ) + nor(FDC ) + (1 nor(OET )) + nor(AEF ))

Afterwards, we can evaluate the overall performance for all the

(MFVi,j , i refers to the number of features and j means the number

of the algorithms and different number of features and test cases

Moreover, Mean Fitness Value for Features (MFVFi ) is dened

where 10 i 1000 with an increment of 10, i.e., a set of hundred

where 5 i 45 with an increment of 10, i.e., a set of ve MFV_TC

5.2. Statistical tests

5.3. Experiment execution

Objectives with the given thresholds

WBGA (WBGA_W1 and WBGA_W2 ), the results do not stay stable,

WBGA_W1 vs. Thresholds

All p-values less than 0.05 are identied as bold.

WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA_W2

WBGA_W1 vs. WBGA_W2

All p-values less than 0.05 are identied as bold.

pair of weight-based GAs as shown in Tables 6 and 7. In total, we

WBGA_W1 vs. WBGA_W2

All p-values less than 0.05 are identied as bold.

Pair of algorithms (A vs. B)

WBGA_W1 vs. WBGA_W2

Objectives with the given thresholds

signicantly less probability to obtain better results than the given

RWGA vs. Thresholds

All p-values less than 0.05 are identied as bold.