Professional Documents
Culture Documents
Thesis submitted to
By
LAXMINARAYAN SAHOO
LAXMINARAYAN SAHOO
2012
Dedicated
to
my beloved and respected teachers
Dr Asoke Kumar Bhunia and Dr Dilip Roy
DECLARATION
Kumar Bhunia and Professor Dilip Roy. I further declare that the work embodied in
this thesis has not been submitted previously, in whole or in part, to any University
Date:
THE UNIVERSITY OF BURDWAN Professor Dilip Roy
GOLAPBAG, BURDWAN – 713 104 E-mail: dr.diliproy@gmail.com
Dr Asoke Kumar Bhunia
WEST BENGAL, INDIA
Associate Professor
E-mail: math_akbhunia@buruniv.ac.in
CERTIFICATE
In our opinion, this thesis is of the standard required for the award of the
degree of “Doctor of Philosophy”.
hilosophy”.
The research works, embodied in this thesis, have not been submitted to
any University or Institution for the award of any degree or diploma.
-------------------------------------- --------------------------------------
(Dr Asoke Kumar Bhunia) (Professor Dilip Roy)
Department of Mathematics Centre for Management Studies
The University of Burdwan The University of Burdwan
Burdwan-713104 Burdwan-713104
India India
Acknowledgements
University of Burdwan and Professor Dilip Roy, Centre for Management Studies, The
not only for their scholarly guidance and encouragement, but also for their endless
and Raniganj Girls’ College for providing me the opportunity to carry out this
research work and for extending whole hearted cooperation and support.
from the Head of the Department and all other faculty members and staff of the
research work. I would like to extend my thanks especially to Professor Gora Chand
College, Dr Ranjan Kumar Gupta, West Bengal State University, Dr Jayanta Majumdar,
Mr Sanat Mahato, Mr Avijit Duary, Mr Debkumar Pal, Mr Amiya Biswas, Mr Akbar Ali
Shaikh and Mr Subhra Sankha Samanta for their constant help and suggestions. I also
thank Mr Biswajit Ta and Mr Tarun Das for helping me a lot during my research
work.
memorable.
including my two little lovely and sweet nieces, Senjuti and Sangbrita. I could not
Most specially, I thank my wife, Anima, for her constant inspiration and
Finally, I want to acknowledge the help, support and love of my elder brother,
In fine, I am solely responsible for any errors and omissions in this thesis.
Date:
Publications
The thesis includes the following published works and a few works communicated
for publication:
Published/Accepted papers
1. Genetic algorithm based multi-objective reliability optimization in interval
environment, Computers & Industrial Engineering, 62, 152-160, 2012.
Communicated papers
2. Reliability optimization under high and low level redundancies via genetic
algorithm for imprecise parametric values, Computers & Structures, Elsevier.
Contents
Certificate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .v
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .vii
Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xv
Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xvi
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1
4.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Bibliography. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .154
Notations used in the Thesis
components
R j ( x j ) = 1 − (1 − rj )
xj The reliability of j -th subsystem
x ( x1 , x2 ,..., xn )
( x, R ) ( x1 , x2 ,..., xq , Rq +1 ,..., Rn )
C w ( x, R ) Weighted cost
t Mean time-to-failure
βi = [ βiL , βiR ] Interval valued Weibull shape parameter for i-th subsystem
[ βiL , βiR ]
ri (t ) = [ riL (t ), riR (t )] e−[αiL ,αiR ]t , i = 1, 2, ⋅⋅⋅, n
subsystem
[ qiL , qiR ] = 1 − [ riL , riR ] Unreliability of i-th component
γi Level of significant
N ( µbi , σ b2i ) Normal distribution with mean µbi and variance σ b2i
Notations… xiv
LN ( µbi , σ b2i ) Log normal distribution with mean µbi and variance σ b2i
zi* = [ ziL* , ziR* ] i-th component of interval valued ideal objective vector
zi** = [ ziL
** **
, ziR ] i-th component of interval valued utopian objective vector
S Feasible region
Table 5.1: Values of α i and βi (i = 1, 2, 3, 4,5) for four different cases. . . . . . . . . . . .101
Figure 3.6: P_size vs. centre of the objective function for Example 1. . . . . . . . . . . . . .74
Figure 3.7: P_cross vs. centre of the objective function for Example 1. . . . . . . . . . . . 74
Figure 3.8: P_mute vs. centre of the objective function for Example 1. . . . . . . . . . . . 75
Figure 4.5: P_size vs. interval valued system reliability for Example 1. . . . . . . . . . . .90
Figure 4.6: Max_gen vs. interval valued system reliability for Example 1. . . . . . . . . 90
Figure 4.7: P_cross vs. interval valued system reliability for Example 1. . . . . . . . . . .91
Figure 4.8: P_mute vs. interval valued system reliability for Example 1. . . . . . . . . . .91
Figure 6.2: P_size vs. interval valued system reliability for series system. . . . . . . .113
Figure 6.3: P_cross vs. interval valued system reliability for series system . . . . . .114
Figure 6.4: P_mute vs. interval valued system reliability for series system . . . . . .114
Figure 6.5: Max_gen vs. interval valued system reliability for series system. . . . .114
Figure 8.2: P_size vs. interval valued system reliability for MOOP. . . . . . . . . . . . . . .148
Figure 8.3: P_cross vs. interval valued system reliability for MOOP. . . . . . . . . . . . . 148
Figure 8.4: P_mute vs. interval valued system reliability for MOOP. . . . . . . . . . . . . 148
Acronyms used in the Thesis
GA Genetic Algorithm
PC Personal Computer
• General Introduction
The subject “Reliability Optimization” appeared in the literature in due late 1940s
and was first applied to communication and transportation systems. Most of the
operating system. One of the goals of the reliability engineer is to find the best way to
increase the system reliability. The reliability of a system can be defined as the
point of time (i.e., mission time) under stated conditions. As systems are becoming
more complex, the consequences of their unreliable behavior have become severe in
terms of cost, effort and so on. The interests in accessing the system reliability and
the need to improve the reliability of products and system have become more and
more important.
increase the system reliability. This can be done by different ways. Some of these are
as follows:
failure occurs.
(v) Using preventive maintenance such that components are replaced by new ones
To improve the system reliability, implementation of the above steps will normally
this problem is to find the number of redundant components that maximizes the
system reliability under several resource constraints. This problem is one of the most
popular ones in reliability optimization since 1950s because of its potentiality for
many cases, redundancy is difficult to add to real systems due to technical limitations
and relatively large quantities of resources, such as weight, volume and cost that are
required.
expansive hardware and software resources and provide the access of main systems
from different locations. These problems have many applications in the areas of
gas sewer networks. During the designing of network systems, one of the important
steps is to find the best layout of components to optimize some performance criteria,
integrated circuits and nano technology, however, have revived the importance of
techniques have been studied. These approaches are mainly based on adding
memory integrated circuits and VLSI, which includes internal memory blocks,
currently use a hierarchical redundancy scheme to increase the yield and reliability
of the chip.
due to several resource constraints arising out of the size, cost and quantities of
problem. To solve this type of problem, several researchers have proposed different
approaches. In their works, the reliabilities of the system components are assumed to
be known at a fixed positive level, which lies between zero and one. However in real-
life situations, the reliabilities of these individual components may fluctuate due to
components with exactly identical reliabilities. Moreover the human factor, improper
storage facilities and other environmental factors may affect the reliabilities of the
positive imprecise numbers between zero and one instead of fixed real numbers. To
define the problem associated with such imprecise numbers, generally different
constraints and goals are considered as fuzzy sets with known membership
viewed as fuzzy sets and others as random variables. However, to select the
to overcome the difficulties arisen in the selection of those, the imprecise numbers
a well known powerful computerized heuristic search and optimization method, viz.
(depending on the evolution principle “Survival of the fittest”) and natural genetics. It
each iteration (which is called generation), three basic genetic operations, viz.
follows:
“Reliability is the probability that a system will perform satisfactorily for at least a
intended purpose adequately for the period of time intended under the operating
conditions encountered. The reliability is the probability with which the devices will
not fail to perform a required operation for certain duration of time. Such problem is
known as the problem of survival. This definition brings into the focus of four
However, in practice, even the best design manufacturing and maintenance efforts do
measure of how well a system meets its design objective and it is usually expressed
up into down to sub systems and elements whose individual reliability factors can be
Introduction 7
elements are connected to constitute the given system. The combinatorial rules are
can be hardware or human or even software. If some of the components are software
The series configuration is the simplest and perhaps one of the most common
structures. In this configuration, all the components must be operating to ensure the
system operation. In other words, the system fails when any one of the components
fails.
A parallel system is a system that is not considered to have failed unless all
system.
and so on. This system has a non-linear and non-separable structure and consists of
there exist combinations of components which are connected neither in a series nor
In non-series systems, it is not necessary that all components operate to make the
system operational. In such systems, we can also find subsets of components such
that the failure of all components in the subset leads to the system failure
irrespective of the states of the other components. The theory of coherent systems
deals with the deterministic functional relationship between the system and its
components. Such a relationship is useful for finding the reliability of large and
complex/complicated systems.
there is no doubt that these technological systems have improved the productivity,
health and affluence of our society. However, this increasing dependence on modern
system reliability plays an important role. The reliability of any system is very
important to manufacturers, designers and also to the users. During the design phase
reliability of that product. They desire the larger reliability of their products which
raise the production cost of the items. In such a case, there arises a question as to
how to meet the goal for the system reliability. As a result, the increase in the
production cost has negative effects on the user’s budget. Therefore, the design
cost. In this connection, a widely known method for improving the system reliability
system using components with known cost, reliability, weight and other attributes,
cost and/or minimum targeted system reliability level. The corresponding problem is
known as the reliability redundancy allocation problem. The primary objective of the
three categories according to the types of decision variables. These are reliability
component reliabilities are the only variables, then the problem is called reliability
allocation. If the number of redundant components is the only variable, then the
problem is called redundancy allocation problem (RAP). On the other hand, if both
the component reliabilities and redundancies are variables of the problem then the
problems, one may refer to the works of Allella, Chiodo and Lauria (2005), Yalaoui,
Chatelet and Chu (2005) and Salzar, Rocco and Galvan (2006). Researchers like Kim
and Yum (1993), Coit and Smith (1996, 1998), Prasad and Kuo (2000), Liang and
Smith (2004), Ramirez-Marquez and Coit (2004), Yun and Kim (2004), Nourelfath
and Nash (2005), You and Chen (2005), Agarwal and Gupta (2006), Coit and Konak
(2006), Ha and Kuo (2006b), Tian and Zuo (2006), Liang and Chen (2007), Nash,
Introduction 11
Nourelfath and Ait-Kadi (2007), Onishi, Kimura, James and Nakagawa (2007), Zhao,
Liu and Dao (2007) and others have solved redundancy allocation problem. Also,
Federowicz and Mazumdar (1968), Tillman, Hwang and Kuo (1977b), Misra and
Sharma (1991), Dhingra (1992), Painton and Campbell (1995), Ha and Kuo (2005,
2006a), Chen (2006), Kim, Bae and Park (2006) and others have solved the reliability
redundancy [Gordon (1957), Messinger and Shooman (1970), Misra (1975), Sakawa
(1978a, 1981a), Zhao and Liu (2003), Yu, Yalaoui, Chatelet and Chu (2007)], multi-
state system reliability [Boland and EL-Neweihi (1995), Prasad and Kuo (2000),
Ramirez-Marquez and Coit (2004), Meziane, Massim, Zeblah, Ghoraf and Rahil
(2005), Tian, Levitin and Zuo (2009) and Li, Chen, Yi and Tao (2010)] and
modular/multi-level redundancy [Yun and Kim (2004) and Yun, Song and
Kim(2007)].
techniques etc. Dynamic programming, branch and bound, cutting plane technique,
implicit enumeration search technique are exact methods which provide exact
determine the optimal solutions is presented in Tillman, Hwang and Kuo (1977a,
1980), Sakawa (1978b, 1981b), Kuo, Prasad, Tillman and Hwang (2001) and Kuo and
Wan (2007a, 2007b). On the other hand, heuristic, meta- heuristic and hybrid
They can provide optimal or near optimal solution in reasonable computational time.
Genetic algorithm, simulated annealing, tabu search, ant colony optimization and
particle swarm optimization are some of the approaches in those categories. For
detailed discussion, one may refer to the works of Kuo, Hwang and Tillman (1978)
Coit and Smith (1996), Hansen and Lih (1996), Ravi, Murty and Reddy (1997), Zhao
and Song (2003), Liang and Smith (2004), Coelho (2009a) and others.
Bellman (1957) and Bellman and Dreyfus (1958, 1962) used dynamic
their works, the problem was to identify the optimal levels of redundancy for only
In the year 1968, Fyffe, Hines and Lee (1968) considered a system having 14
subsystems with both cost and weight constraints and solved the corresponding
for each subsystem there are three or four different choices of components each with
different reliability, weight and cost. They used Lagrange’s multiplier technique to
programming. Their algorithm was tested for 33 different Fyffe’s problems of which
i.e., integer programming approach. Ghare and Taylor (1969) first used the branch
and bound method to maximize the system reliability under given non-linear but
separable constraints. Bulfin and Liu (1985) formulated the problem as a knapsack
it. Nakagawa and Miyazaki (1981) investigated the Fyffe problem and formulated 33
variances of the problem as integer programming problem. Bulfin and Liu (1985)
programming problems like those of Ghare and Taylor (1969). The problem was
reliability, cost and weight and also solved by integer programming by Gen, Ida,
proposed by researchers, like, Kuo, Lin, Xu and Zhang (1987), Hikita, Nakagawa and
Narihisa (1992), Sung and Cho (1999), Mettas (2000), Coit and Smith (1996, 2002),
Sun and Li (2002), Ha and Kuo (2006b), Liang and Chen (2007), Ramirez-Marquez
problems have been received warm reception among the researchers. In this
connection one may refer to the work of Painton and Campbell (1994, 1995). They
improvements and the level of effort spent on those improvements to maximize one
constraints (e.g., cost) in the presence of uncertainty. In the year 1994, a redundancy
allocation problem with several failure modes was solved by Ida, Gen and Yokota
(1994) with the help of GA. Coit and Smith (1996) have solved a redundancy
components in which each subsystem is a k-out-of-n: G system. In the year 1998, Coit
Studies on Reliability Optimization Problems by Genetic Algorithm 14
and Smith (1998) used GA-based approach to solve the redundancy allocation
percentile of the system time to failure distribution. Yun and Kim (2004) and Yun,
there occurs the simultaneous optimization of more than one objective function.
system reliability, system cost, volume and weight. For this reason multi-objective
optimization problem attracts a lot of attention from the researchers. The objective
of this problem is to maximize the system reliability and minimize the system cost,
volume and weight. A Pareto optimal set, which includes all of the best possible
solutions between the given objectives than a single objective, is usually identified
for multi-objective optimization problems. Dhingra (1992), Rao and Dhingra (1992)
used goal programming formulation and the goal attainment method to generate
Pareto optimal solutions. Ravi, Reddy and Zimmermann (2000) presented fuzzy
the fuzzy goals. Busacca, Marseguerra and Zio (2001) developed a multi-objective GA
target as a separate objective. Sasaki and Gen (2003a, 2003b) solved multi-objective
both objectives and constraints. Elegbede and Adjallah (2003) solved multi-objective
Introduction 15
Tian and Zuo (2006) and Salzar, Rocco and Galvan (2006) have used a genetic
reliability design problems using feature models. In the year 2009, Okasha and
using genetic algorithm. Li, Liao and Coit (2009) have used multiple objective
problem.
mentioned earlier this problem came into the existence in the late 1940s and was
first applied to communication and transport system. After that a lot of works has
works, the reliability of a component was considered as precise value i.e., fixed lying
between zero and one. However, due to some factors mentioned in Section 1.1, it may
not be fixed though it may vary between zero and one. So, to represent the same,
significant. To the best of our knowledge, very few works have been done
considering interval valued reliabilities. Even today, there is a lot of scope to work in
this area considering interval valued reliabilities of components. The detailed scheme
Studies on Reliability Optimization Problems by Genetic Algorithm 16
of works along with the works presented in this thesis and also the further scope of
It may be noted from Figure 1.1 that the objectives of the thesis is
components.
(iii) to solve the problems mentioned in (i) and (ii) by real coded genetic
In this thesis, some reliability optimization problems have been formulated and
solved in interval environment with the help of interval mathematics, our proposed
interval order relations and real coded genetic algorithm. The entire thesis has been
Chapter 1 Introduction
Environment
Valued Parameters
Studies on Reliability Optimization Problems by Genetic Algorithm 18
Stochastic Domain
interval order relations and real coded genetic algorithm. In this chapter, we have
also proposed new definition of interval power of an interval and new order
systems considering the reliability of each component as interval valued number. For
of interval arithmetic and interval order relations. Then the problems have been
function techniques. To solve these problems, two different real coded genetic
algorithms (GAs) for interval valued fitness function with tournament selection,
whole arithmetical crossover and boundary mutation for floating point variables,
intermediate crossover and uniform mutation for integer variables and elitism with
size one have been developed. To illustrate the models, some numerical examples
have been solved and the results have been compared. As a special case, taking lower
Introduction 19
and upper bounds of the interval valued reliabilities of component as same, the
corresponding problems have been solved and the results have been compared with
the results available in the existing literature. Finally, to study the stability of the
proposed GAs with respect to the different GA parameters (like, population size,
crossover and mutation rates), sensitivity analyses have been shown graphically.
that maximizes the overall system reliability subject to the given resource
constraints and also minimizes the overall system cost subject to the given resources
component is interval valued and the cost coefficients as well as the amount of
resources are imprecise and interval valued. These types of problems have been
In this work, we have formulated two types of redundancy, viz. component level
redundancy known as low-level redundancy and the system level redundancy known
unconstrained problem using penalty function technique and solved using genetic
algorithm. Finally, two numerical examples (one for low-level redundancy and
another for high-level redundancy) have been presented and solved and the
Weibull distribution with imprecise parameters. In the earlier work, either both the
reality, both the parameters may vary due to some factors and it is sensible to treat
and solved by genetic algorithm. Finally, a numerical example has been solved for
optimization problem in the series system. This problem can be formulated as a non-
under chance constraints due to resources. In this chapter, we have formulated the
solve this problem, we have developed a real coded genetic algorithm (GA) for
neighborhood mutation. To illustrate the model, two numerical examples have been
considered and solved by our developed GA. Finally to study the stability of our
the stochastic domain with respect of resource constraints and the concept of
interval valued parameters has been integrated with the stochastic setup so as to
increase the applicability of the resultant solutions. In particular, the five-link bridge
network system has been studied under a normal setup with Genetic Algorithm as
by maximizing the system reliability and minimizing the system cost under several
have been formulated in interval environment with the help of interval mathematics
and our newly proposed order relations of interval valued numbers. Then these
optimization problems have been solved by advanced genetic algorithm and the
concept of Pareto optimality. Finally, for the purpose of illustration and comparison,
In Chapter 9, general concluding remarks drawn from our studies and further
• Metric Space
• Genetic Algorithm
During the last few decades, several researchers formulated and solved either single
problems, they proposed different techniques. In this connection, one may refer to
the works of Tillman, Hwang and Kuo (1977a, 1977b and 1980), Nakagawa,
Nakashima and Hattori (1978), Misra and Sharma (1991), Chern (1992), Ohtagaki,
Nakagawa, Iwasaki, and Narihisa (1995), Kuo, Prasad, Tillman and Hwang (2001),
Sun and Li (2002), Gen and Yun (2006), Ha and Kuo (2006b), Coelho (2009a, 2009b)
optimization have been taken to be precise values. This means that every probability
involved is perfectly determinable. In this case, it is usually assumed that there exist
some complete probabilistic information about the system and the component
behavior. However, in real-life situations, there are not sufficient statistical data
available in most of the cases where the system is either new or it exists only as a
project. It is not always possible to observe the stability from the statistical point of
view. This means that only some partial information about the system components
are known. In these cases, parameters are said to be imprecise. To tackle the problem
approaches are applied and the corresponding problems are converted into
deterministic problems for solving them. In the stochastic approach, the parameters
fuzzy approach, the parameters, constraints and goals are considered as fuzzy sets
with known membership functions or fuzzy numbers. On the other hand, in the
Studies on Reliability Optimization Problems by Genetic Algorithm 24
maker to specify the appropriate membership function for fuzzy approach and
approach. So, to avoid these difficulties for handling the imprecise parameters by
aR are the left and right limits respectively and » is the set of all real numbers.
Actually, every real number can be treated as an interval, such as for all x ∈ » , x can
A + B = [aL + bL , aR + bR ] , A + B = 〈 ac + bc , aw + bw 〉
A − B = [aL − bR , aR − bL ]
[λ aL , λ aR ] for λ ≥ 0,
λA =
[λ aR , λ aL ] for λ < 0,
or, λ A = λ 〈 ac , aw 〉 = 〈λ ac , λ aw 〉 .
aL + aR
m( A) =
2
B 1 1 1
= B × = [bL , bR ] × [ , ], provided 0 ∉ [ aL , aR ] .
A A aR aL
The above definitions are given in the books written by Moore (1979) and Hansen
interval is as follows:
[1,1] if n = 0
n n
[aL , aR ] if aL ≥ 0 or if n is odd
An = n n
[aR , aL ] if aR ≤ 0 and n is even
n n
[0, max(aL , aR )] if aL ≤ 0 ≤ aR and n > 0 is even
Studies on Reliability Optimization Problems by Genetic Algorithm 26
[ n aL , n aR ] if aL ≥ 0 or if n is odd
1 1
( A) n = [ aL , aR ]n = n [aL , aR ] = [0, n aR ] if aL ≤ 0, aR ≥ 0 and n is even
ϕ if aR < 0 and n is even
Again applying the definitions of power and different roots of an interval, the rational
p 1 p
q p q p q
( A) = ( A ) or equivalently, ( A) = exp log A , provided aL > 0 .
q
Let A = [aL , aR ] and B = [bL , bR ] . A complex interval z is identified with the interval
intervals like, addition, subtraction, division and multiplication be the same as the
f ( A) = f ([aL , aR ]) = [ f ( aL ), f (aR )] .
Solution methodologies 27
Using the above definitions the exponential and logarithmic function can be
complicated.
π
−1 if ∃ k ∈ » : 2kπ − ∈[aL , aR ]
where bL = 2
min{sin(aL ), sin(aR )} otherwise
π
1 if ∃ k ∈ » : 2kπ + ∈[aL , aR ]
and bR = 2
max{sin(aL ), sin(aR )} otherwise
1 if ∃ k ∈ » : 2kπ ∈[aL , aR ]
and bR =
max{cos(aL ), cos(aR )} otherwise
b b b
functions.
Till now, none has developed the interval power of an interval number. In this thesis,
in Chapter 5 and Chapter 8, the interval power of an interval numbers occurs in the
formulation of the optimization problems. For this purpose, we have introduced the
[exp(u ), exp(v)] if aL ≥ 0
=
a complex interval if aL < 0
= [ aR , aL ][bL ,bR ] + cos[(2k + 1)π bL , (2k + 1)π bR ] + i sin[(2k + 1)π bL , (2k + 1)π bR ]
if aL , aR ≥ 0, k = 0,1, 2,3, ⋅⋅⋅
The mean, variance and standard deviation of n interval numbers are defined as
follows:
number. Then mean ( x ) , variance [Var( x)] and standard deviation (σ x ) of these
1 n 1 n
x = [ xL , xR ] = ∑ xiL , ∑ xiR ,
n i =1 n i =1
2
1 n 1 n 1 n
Var(x) = [σ L2 , σ R2 ] = ∑ [ xiL − ∑ xiR , xiR − ∑ xiL ] ,
n i =1 n i =1 n i =1
and σ x = [σ L , σ R ] = Var(x) .
For obtaining the optimum solution in solving the optimization problems with
interval valued objectives we need to define the order relations of interval numbers.
Let A = [aL , aR ] and B = [bL , bR ] be two unequal intervals. Then these two intervals
Type-3: One of the intervals contains the other one [see Figure 2.3].
B A
A B
aL aR bL bR bL bR aL aR
B B
A A
aL bL aR bR bL aL bR aR
B B
A
A
aL bL bR aR bL aL aR bR
It is to be noted that both the intervals A = [aL , aR ] and B = [bL , bR ] will be equal in
In this area, very few researchers defined the order relations of interval
valued numbers. Moore (1979) first proposed two order relations of interval
numbers.
For any two intervals A = [aL , aR ] and B = [bL , bR ] , Moore (1979) first gave the
However, these two order relations cannot order two partially or fully overlapping
intervals. Then Ishibuchi and Tanaka (1990) defined the order relations for
These order relations are reflexive, transitive and anti-symmetric i.e., these are
partial order. From these definitions it is clear that, for minimization problem, a
and Tanaka (1990), Chanas and Kuchta (1996) proposed the concept of t0t1 − cut of
Let A = [aL , aR ] be any interval and t0 and t1 be any two fixed numbers such
According to Chanas and Kuchta (1996), the order relations for the intervals
(ii) A ≤cw /[t0 ,t1 ] B iff A /[t0 ,t1 ] ≤cw B /[t0 ,t1 ]
After Chanas and Kuchta (1996), Kundu (1997) first noticed that the interval ranking
methods discussed earlier could not find the measure ‘How much larger the interval
introducing the ‘fuzzy leftness relation’. For the intervals A and B , let a ∈ A and
Then A is left to B if Left( A, B ) = max{0, P ( a < b) − P (a > b)} > 0 and A is right to B
if Right( A, B ) = max{0, P ( a > b) − P ( a < b)} > 0 , where P ( a < b) denotes the probability
In the year 2000, two other approaches of ranking of two intervals were given
by Sengupta and Pal (2000). In the first approach, they defined order relations with
Studies on Reliability Optimization Problems by Genetic Algorithm 32
bc − ac
χ : I × I → [0, ∞) for the intervals A and B as χ ( A, B) = , where bw + aw ≠ 0 .
bw + aw
A cannot be accepted as smaller. If 0<χ ( A, B ) < 1 , A can be accepted with the grade of
bc − ac
acceptability . The interval A is accepted fully if χ ( A, B) = 1 .
bw + aw
According to them, the acceptability index is only a value based ranking index
and it can be applied partially to select the best alternative from the pessimistic point
of view of the decision-maker. So, only the optimistic decision-maker can use it
completely. In another approach, Sengupta and Pal (2000) introduced the fuzzy
preference ordering for the ranking of a pair of interval numbers on the real line with
was described for maximizing the profit interval. However, this method is equally
applicable to minimize the cost/time intervals also. In this definition, they assumed
that two intervals A and B are profit intervals and the problem is to find the
maximum profit interval from among them. In this approach, they considered the
of B in comparison to A ”.
1 if bc = ac
b −a −b
µ ( B, A) = max 0, c L w if aL + bw ≤ bc ≤ ac
ac − aL − bw
0 otherwise
This non-linear membership function lies in the interval [0, 1]. When the values of
Solution methodologies 33
this membership function lies within the interval [0.333, 0.666], this definitions fails
Mahato and Bhunia (2006) proposed the revised definitions of order relations
between interval costs/times for minimization problems and interval profits for
makers’ risk taking attitude. In this case, they considered two types of decision-
problems and highest value for maximization problems ignoring the uncertainty.
Definition: For minimization problems, the order relation ≤omin between the
A ≤omin B ⇔ aL ≤ bL
This implies that A is superior to (i.e., better than) B and A is accepted but B is not
Definition: For maximization problems, the order relation ≥omax between the
A ≥omax B ⇔ aR ≥ bR
This implies that A is superior to B and the optimistic decision-maker accepts the
profit interval A. Here also, the order relation ≥omax is not symmetric.
Definition: For minimization problems, the order relation < pmin between the
(i) A < pmin B ⇔ ac < bc for Type-1 and Type-2 intervals and
However, for Type-3 intervals with (ac < bc ) ∧ (aw > bw ) , a pessimistic decision
Definition: For maximization problems the order relation > pmax between the
(i) A > pmax B ⇔ ac > bc , for Type -1 and Type-2 intervals and
Again, for Type-3 intervals with (ac > bc ) ∧ (aw > bw ) , a pessimistic decision cannot be
To overcome this situation, we have proposed two new definitions of order relations
Solution methodologies 35
Definition: The order relation > max between the intervals A = [aL , aR ] = ac , aw
(i) A > max B ⇔ ac > bc for Type-1 and Type-2 intervals and
According to this definition, the interval A is accepted for maximization case. Clearly,
this order relation > max is reflexive and transitive but not symmetric.
problems
According to this definition, the interval A is accepted for minimization case. Clearly,
the order relation <min is reflexive and transitive but not symmetric.
It is to be noted that the definitions given by Mahato and Bhunia (2006) and our
proposed definitions give the same results. So these two definitions are equivalent.
The term metric is derived from the term metor (measure). The concept of a metric
(i) d ( x, y ) ≥ 0, ∀ x, y ∈ X
(ii) d ( x, y ) = 0 ⇔ x = y, ∀ x, y ∈ X
(iii) d ( x, y ) = d ( y, x), ∀ x, y ∈ X
(iv) d ( x, y ) ≤ d ( x, z ) + d ( z, y ), ∀ x, y , z ∈ X
metric d is also called a distance function. The real number d ( x, y ) is called the
Here, we shall give some metric spaces which are used in this thesis.
(i) Let X = » n and suppose x = (ξ1 , ξ 2 ,..., ξ n ) and y = (η1 ,η2 ,...,η n ) be any two points in
{ p
d p ( x, y ) = ξ1 − η1 + ξ 2 − η2
p
+ ⋅⋅⋅ + ξ n − η n }
p p
, 1≤ p < ∞
Then, d p for each 1 ≤ p < ∞ and d ∞ are metrics on the same underlying set X = » n .
(iii)Let X = l p ,1 ≤ p < ∞ , be the set of all sequences x = {ξi } of real scalars such
1
∞ n p p
that ∑ ξi < ∞ . Define the mapping d : X × X → » by d ( x, y ) = ∑ ξi − ηi ,
p
i =1 i =1
the evolutionary theory of Charles Darwin “survival of the fittest” and natural
Solution methodologies 37
transportation and assignment problems, job scheduling, inventory control and other
selection. The concept of GA was first introduced by Prof. John Holland of the
University of Michigan, Ann Arbor. He is considered to be the father of GA. His idea of
genetic algorithm was first used to solve optimization problem by De-Jang (1975).
Thereafter, a researcher has contributed much to the major development of this field.
Most of the initial research work in this field can be found in several International
extensions along with related topics, can be found in the books on GA [Holland
(1975), Goldberg (1989), Davis (1991), Michalewicz (1996), Mitchell (1996), Gen
Keeping the imitation of natural evolution as the foundation, genetic algorithm can
solve. This algorithm starts with an initial population of probable solutions, called
form of coding as a chromosome. These chromosomes are evaluated for their fitness.
Studies on Reliability Optimization Problems by Genetic Algorithm 38
operations, like crossover and mutation. The crossover operation is applied to create
offspring from a pair of selected chromosomes. The mutation operation is used for a
solutions to the problem under investigation. This process is applied iteratively until
2.3.1.1 GA Terminology
problem
chromosome.
the problem.
specified generations.
positive number.
number of generations).
Studies on Reliability Optimization Problems by Genetic Algorithm 40
2.3.1.2 Algorithm
the variables.
population at t - th generation].
next step.
Step-12 : Compare the best chromosome of P (t ) and P (t − 1) and store better one.
Step-13 : Go to step-6.
Step-14 : Print the best chromosome (which is the solution of the optimization
problem).
Step-15 : End.
Solution methodologies 41
2.3.1.3 GA Parameters
There are several GA parameters, viz. population size (p_size), maximum number of
mutation rate i.e., the probability of mutation (p_mute). There is no hard and fast rule
for selecting the population size for GA, how large it should be. The population size is
problem dependent and will need to increase with the dimensions of the problem. If
the population size is very large, storing of data in intermediate steps of GA may arise
some difficulties at the time of execution. When the population size is very small,
some genetic operators do not work properly. However, population size is restricted
by both time complexity and space complexity. Regarding the maximum number of
generations, there is no clear indication for considering this value. It varies from
solution has converged. From natural genetics, it is obvious that the rate of crossover
is always greater than that of rate of mutation. Generally, the crossover rate varies
from 0.60 to 0.95 whereas the mutation rate varies from 0.05 to 0.20. Sometimes
Studies on Reliability Optimization Problems by Genetic Algorithm 42
chromosome.
GA for solving the optimization problem and users of GA face a hard situation how to
representations, like, binary, real, octal, hexadecimal coding, available in the existing
literature. Among different representations, mainly binary coding and real coding
representation, binary sub strings of each variable with the desired precision are
will be large. In this case, genetic algorithm would perform poorly. To overcome
these difficulties, real numbers are used to represent the chromosomes in GAs. In
both numbers, with the same component as the vector of decision variables of the
the real design space and moreover, the string length of each chromosome is the
( k = 1, 2,..., p _ size ) is an ordered list of n genes as vk = {vk1 , vk 2 ,..., vki ,..., vkn } .
Solution methodologies 43
2.3.1.5 Initialization
that will take part in the artificial genetics. To initialize the population, first of all we
have to find the independent variables and their bounds for the given problem. Then
every component for each chromosome is randomly generated within the bounds of
the corresponding decision variable. There are several procedures for selecting a
random number of either integer type or float point type. In our whole work, we have
either x = a + g or, x = b − g
On the other hand, to generate a floating point random number, we have used
Uniform Distribution
The initialization procedure produces p_size chromosomes where p_size denotes the
Step-2: Assign xi = li + r (ui − li ) ), where li and ui are lower and upper bounds of xi .
Step-3: Repeat the steps 1 and 2 for n times and produce a vector ( x1 , x2 ,..., xn ).
Step-4: Repeat the steps 1, 2 and 3 for p_size times and produce p_size initial feasible
solutions.
Studies on Reliability Optimization Problems by Genetic Algorithm 44
Step-5: Stop.
It may be noted that a random value can be selected alternatively from the discrete
set of values {0, 0.1, 0.2, 0.3,…, 0.9}. Generally, this process is used to find the low
random number r is selected from either discrete set of values or Step-1 of earlier
Evaluation/fitness function plays an important role in GA. This role is same for
good they are. Therefore, we have to calculate the fitness value for each chromosome.
2.3.1.7 Selection
The selection operator which is the first operator in artificial genetics plays an
interesting role in GA. This selection process is based on the Darwin’s principle on
natural evolution “survival of the fittest”. The primary objective of this process is to
the fitness value of each chromosome and eliminate the rest of the
process.
problems. As a result, for solving those problems we have used only the tournament
selection with size two. In this selection, two individuals in the population are
selected based on their fitness. The following assumptions for this selection
(i) when both the individuals/chromosomes are feasible then the one with better
(ii) when one individual/chromosome is feasible and another is infeasible then the
(iii) when both the individuals/chromosomes are infeasible with unequal constraint
(iv) when both the individuals/chromosomes are infeasible with equal constraint
2.3.1.8 Crossover
The exploration and exploitation of the solution space can be made possible by
process, other genetic operators, like crossover and mutation are applied to the
parent solutions. It operates on two or more parent solutions at a time and produces
Studies on Reliability Optimization Problems by Genetic Algorithm 46
offspring for next generation. In this operation, expected [ p _ cross * p _ size] number
of chromosomes will take part (* and [ ] denote the product and the integral value
respectively).
Step-2: Select two parent chromosomes vk and vi randomly from the population.
Step-3: Compute the components vkj and vij ( j = 1, 2,..., n) of two offspring from the
N
Step-4: Repeat Step-2 and Step-3 for times.
2
In case of whole arithmetic crossover, components vkj and vij ( j = 1, 2,..., n) of two
The aim of mutation operator is to introduce the random variations into the
population and is used to prevent the search process from converging to the local
optima. This operator helps to regain the information lost in earlier generations and
is responsible for fine tuning capabilities of the system and is applied to a single
individual only. Usually, its rate is very low; because otherwise it would defeat the
order building being generated through the selection and crossover operations.
Step-3: Select a particular gene vik (k = 1, 2,..., n) on chromosome vi for mutation and
Step-4: Create new gene vik′ ( k = 1, 2,..., n) corresponding to the selected gene vik by
mutation process.
Among these mutation operators, one-neighborhood mutation is new. For the first
time, we have proposed this operator. Basically, this operator is used in GA to mutate
the gene corresponding to integer variables. Other two operators are well known
After mutation process, let vik' (k = 1, 2,..., n) be the mutated gene corresponding to the
2.3.1.9 Elitism
In any generation of GA, sometimes there arises a situation when the best
chromosome may get lost from the population when a new population is created by
or more chromosomes may take part in this operation. This process is named as
elitism.
It
(ii) optimizes the objective function with continuous, discrete, permutation and
mixed variables.
(viii) works not only with the analytical functions, but also works with experimental
data.
(ix) works with a set of solutions instead of single solution in each iteration/
generation.
(x) is able to solve problem with non-convex solution space, where classical
(xi) It also performs well with problems where the objective function is interval
All these advantages make the GAs superior from the classical optimization
types of optimization problems, there are few disadvantages also. These are as
follows:
(i) It is often seen that a genetic algorithm caught in a local optimum, and that all or
most of the population concentrates on a small part of the search space located
(ii) The most difficult and time consuming issue in the successful application of GAs
is an art and the quality of this tuning gently depends on the user-experience as
arises an important question for handling the constraints relating to the problem.
During the past, several methods have been proposed to handle the constraints in
evolutionary algorithms for solving the same problem with fixed objective. These
methods can be classified into several types, viz. penalty function techniques,
methods that preserve the feasibility of solutions, methods that clearly distinguish
between feasible and infeasible solutions. Among these methods, penalty function
technique is very well known and widely applicable. In this technique, the amount of
When the objective function is increased/ decreased with a penalty term multiplied
by so called penalty coefficient there arises a difficulty to select the initial value and
upgrading strategy for the penalty coefficient. To overcome this difficulty, Deb
technique, the worst fitness value of GA for feasible solutions is considered as the
fitness value of infeasible solution without multiplying the penalty coefficient i.e., the
g j ( x) ≤ b j , j = 1, 2,..., m ,
Therefore, according to the PFP technique, the converted problem of problem (2.1) is
as follows:
[0, 0] if x ∈ S
where θ ( x) =
−[ f L ( x), f R ( x)] + min[ f L ( x), f R ( x)] if x ∉ S
and S = {x : g j ( x ) ≤ b j , j = 1, 2,..., m}
function of the worst feasible solution in the population. Alternatively, the problem
may be solved with another fitness function by penalizing a large positive number
(say M which can be written in the interval form as [ M , M ] ) [Gupta, Bhunia and Roy
(2009)]. We denote this penalty as Big-M penalty and its form is as follows:
[0, 0] if x ∈ S
where θ ( x) =
−[ f L ( x), f R ( x)] + [ − M , − M ] if x ∉ S
The above problems (2.2) and (2.3) are non-linear unconstrained optimization
In our work, we have used Big-M penalty technique with the value of M as 99999.
CHAPTER 3
Reliability Redundancy Allocation
Problems in Interval
Environment
• Introduction
• Solution Procedures
• Numerical Examples
• Sensitivity Analysis
• Concluding Remarks
Reliability Redundancy Allocation Problems in Interval Environment 53
3.1 Introduction
productivity, our modern society has become more delicate and vulnerable due to
complicated operations and highly sophisticated management. From any respect, the
important question how to obtain a balance between reliability and other resources
e.g., cost, volume and weight. In the last few decades, several researchers considered
constraints [Tillman, Hwang and Kuo (1977a, 1977b), Nakagawa, Nakashima and
Hattori (1978), Misra and Sharma (1991), Chern (1992), Ohtagaki, Nakagawa,
Iwasaki and Narihisa (1995), Kuo, Prasad, Tillman and Hwang (2001), Sun and Li
(2002), Gen and Yun (2006), Ha and Kuo (2006b) and Coelho (2009a, 2009b)]. To
solve those problems, different techniques have been proposed by the several
may not be fixed. It may vary due to several reasons. There is no technology by which
different components can be produced with exactly identical reliabilities. So, the
number instead of a fixed real number. Studies of the system reliability where the
component reliabilities are imprecise and/or interval valued have already been
Studies on Reliability Optimization Problems by Genetic Algorithm 54
initiated by some authors [Coolen and Newby (1994), Utkin and Gurov (1999, 2001)
and Gupta, Bhunia and Roy (2009)]. To tackle the problem with such imprecise
numbers, generally stochastic, fuzzy and fuzzy-stochastic approaches are applied and
the corresponding problems are converted into deterministic problems for solving
parameters, constraints and goals are considered as fuzzy sets with known
approach, some parameters are viewed as fuzzy sets/fuzzy numbers and others as
appropriate membership function for fuzzy approach and probability distribution for
stochastic approach and both for fuzzy -stochastic approach. So, to avoid these
difficulties for handling the imprecise numbers by different approaches, one may use
most significant representation among others. Due to this representation, the system
approaches for solving reliability optimization problems with the interval objective.
As the objective function of the reliability optimization is interval valued, to solve this
type of problem by GA method, order relations of interval numbers are essential for
selection operation as well as for finding the best chromosome in each generation.
problems with interval coefficients for maximizing the overall system reliability and
system cost under some resource/budget constraints. During the last few years,
several techniques were proposed for solving the constrained optimization problem
with fixed coefficients with the help of GAs [Goldberg (1989), Michalawich (1996)
and Gen and Cheng (2000)]. Among these methods, penalty function techniques are
very popular in solving the same by GAs [Deb (2000), Miettinen, Makela and
Toivanen (2003) and Aggarwal and Gupta (2005)]. This method transforms the
unconstrained one by two different types of penalty function techniques and the
resulting objective function would be interval valued. So, to solve this problem we
have developed two different GAs for integer variables with the same GA operators
like tournament selection, intermediate crossover for integer variables and whole
arithmetical crossover for floating point variables, uniform mutation for integer
variables and boundary mutation for floating point variables and elitism of size one
methods have been illustrated with some numerical examples and to test the
performance of these methods, the results have also been compared. As a special
case considering the lower and upper bounds of interval valued reliabilities of
components as same, the resulting problem becomes identical with the existing
problem available in the literature. Finally, to study the stability of the proposed GAs
with respect to the different GA parameters (like, population size, crossover and
3.2.1 Assumptions
components.
It is well known that a series system (see Figure 3.1) with n independent
components must be operating only if all the components are functioning. In order to
improve the overall reliability of the system; one can use more reliable components.
However, the expenditure and more often the technological limits may prohibit an
limited resource constraints. These constraints may arise out of the size, cost and
optimization problem for such a system for interval valued values of reliability can
be formulated as follows:
q
x x
Problem 1 Maximize [ RSL , RSR ] = ∏ [{1 − (1 − r jL ) j },{1 − (1 − rjR ) j }]
j =1
subject to gi ( x) ≤ bi , i = 1, 2,..., m
where r j = [rjL , r jR ]
Reliability Redundancy Allocation Problems in Interval Environment 57
objective.
1 2 3 n
1 1 1
2 2 2
x1 x2 xq
and so on. For example let us consider a HSP system ( n = 10) shown in the Figure 3.3.
1 2
7
4
3
8 10
5 6
This system has a non-linear and non-separable structure and consists of nested
Problem 2
subject to gi ( x) ≤ bi , i = 1, 2,..., m
When a system can be reduced to series and parallel configurations, there exist
system is also called the bridge system. For example, let us consider a bridge
system ( n = 5) shown in Figure 3.4. This system consists of five subsystems and three
follows:
Reliability Redundancy Allocation Problems in Interval Environment 59
Problem 3
Maximize[ RSL , RSR ] = [ R5 L , R5 R ](1 − [Q1L , Q1R ][Q3 L , Q3 R ])(1 − [Q2 L , Q2 R ][Q4 L , Q4 R ])
subject to gi ( x) ≤ bi , i = 1, 2,..., m
1 2
3 4
1 2
2 3
1 3
n n
as RS = ∑ r j (1 − r )n − j , where r is the component reliability.
j =k j
xj
xj
q
Problem 4 Maximize [ RSL , RSR ] = ∏ ∑ [rL , rR ]l ([1,1] − [rL , rR ]) j
x −l
j =1 l = k l
subject to gi ( x) ≤ bi , i = 1, 2,..., m
Let us consider a network with n subsystems. The goal of the redundancy allocation
overall system reliability subject to the given resource constraints and also to
minimize the overall system cost subject to the given constraint on system reliability.
as follows:
subject to Ci ( x, R ) ≤ bi , i = 1,...., m,
Problem 6 Minimize Cw ( x, R)
subject to RS ( x, R) ≥ R*
In this section we shall discuss the solution procedure for all the problems
mentioned in earlier section i.e., problems (1)-(6). These problems are non-linear
function. Using PFP and Big-M penalty techniques and real coded genetic algorithm
optimization problems.
[0, 0] if x ∈ S
where θ ( x) =
−[ RSL ( x), RSR ( x)] + min[ RSL ( x), RSR ( x)] if x ∉ S
Here the parameter min[ RSL ( x), RSR ( x)] is the value of interval valued objective
function of the worst feasible solution in the population. Alternatively, the problem
may be solved with another fitness function by penalizing a large positive number
(say M which can be written in the interval form as [ M , M ] ) [Gupta, Bhunia and Roy
(2009)]. We denote this penalty as Big-M penalty and its form is as follows:
[0, 0] if x ∈ S
where θ ( x) =
−[ RSL ( x), RSR ( x)] + [− M , − M ] if x ∉ S
The above problems (3.1) and (3.2) are non-linear unconstrained integer
follows:
m
Maximize Rˆ S ( x, R ) = RS ( x, R ) − ∑ [ max(0, Ci ( x, R ) − bi ), max(0, Ci ( x, R ) − bi ) ] + θ ( x, R )
i =1
(3.3)
[0, 0] if ( x, R) ∈ S
where θ ( x, R) =
− RS ( x, R) + min [ RSL ( x, R), RSR ( x, R) ] if ( x, R) ∉ S
Here min [ RSL ( x, R), RSR ( x, R) ] is the value of the interval valued objective function of
Alternatively, the problem may also be solved with another fitness function by
Maximize Rˆ S ( x, R ) = RS ( x, R) + θ ( x, R) (3.4)
[0, 0] if ( x, R) ∈ S
where θ ( x, R) =
− RS ( x, R) + [ − M , − M ] if ( x, R ) ∉ S
m
Minimize Cˆ w ( x, R) = Cw ( x, R) + ∑ max(0, − RSL ( x, R) + R* ) + θ ( x, R) (3.5)
j =1
[0, 0] if ( x, R ) ∈ S
where θ ( x, R) =
−Cw ( x, R) + max {Cw ( x, R)} if ( x, R) ∉ S
{
and S = ( x, R) : − RSL ( x, R) + R* ≤ 0, i = 1, 2,..., m and l ≤ x ≤ u , L ≤ R ≤ U }
where l = (l1 , l2 ,..., lq ) , u = (u1 , u2 ,..., uq ) , x = ( x1 , x2 ,..., xq ) , L = ( Lq +1 , Lq + 2 ,..., Ln ) ,
Here max {Cw ( x, R)} is the value of the interval valued objective function of the worst
Alternatively, the problem may also be solved with another fitness function by
Minimize Cˆ w ( x, R ) = Cw ( x, R ) + θ ( x, R) (3.6)
[0, 0] if ( x, R) ∈ S
where θ ( x, R) =
− Cw ( x, R ) + M if ( x, R) ∉ S
Studies on Reliability Optimization Problems by Genetic Algorithm 64
{
and S = ( x, R) : − RSL ( x, R ) + R* ≤ 0, i = 1, 2..., m and l ≤ x ≤ u , L ≤ R ≤ U }
where l = (l1 , l2 ,..., lq ) , u = (u1 , u2 ,..., uq ) , x = ( x1 , x2 ,..., xq ) , L = ( Lq +1 , Lq + 2 ,..., Ln ) ,
The proposed GAs (viz. PFP-GA and Big-M-GA) have been illustrated for solving
mutation corresponding to the integer variables and whole arithmetic crossover and
purpose we have solved twelve numerical examples. It may be noted that for solving
the said problem with fixed valued reliabilities of components, the reliability of each
component is taken as interval with the same lower and upper bounds of interval.
For each example, 20 independent runs have been performed by both the GAs, of
run in LINUX environment. The computational work has been done on the PC which
has Intel Core-2 duo processor with 2.5 GHz. In this computation, different
population size has been taken for different problems. However, the crossover and
Reliability Redundancy Allocation Problems in Interval Environment 65
mutation rates are taken as 0.95 and 0.15 respectively. The computational results of
different examples have been shown in Tables 3.3, 3.4, 3.5 and 3.6.
5
x x
Maximize [ RSL , RSR ] = ∏ [{1 − (1 − r jL ) j },{1 − (1 − rjR ) j }]
j =1
subject to
5
∑ p j x j2 − P ≤ 0
j =1
5
∑ c j [ x j + exp( x j 4)] − C ≤ 0
j =1
5
∑ w j x j exp( x j 4) − W ≤ 0
j =1
The values of different parameters along with the interval valued reliabilities of
subject to
The values of different parameters along with the interval valued reliabilities of
j rj pj P cj C wj W
1 [0.76, 0.83] 1 7 7
[0.82, 0.87] 2 7 8
2
110 175 200
3 [0.88, 0.93] 3 5 8
4 [0.61, 0.67] 4 9 6
5 [0.70, 0.80] 2 4 9
j rj cj wj lj uj
1 [0.80, 0.84] 8 16 1 4
2 [0.87, 0.90] 4 6 1 5
3 [0.89, 0.93] 2 7 1 6
4 [0.84, 0.86] 2 12 1 7
5 [0.88, 0.90] 1 7 1 5
6 [0.90, 0.95] 6 1 1 5
7 [0.80, 0.85] 2 9 1 3
8 [0.91, 0.95] 8 - 1 3
9 [0.80, 0.83] - - 1 4
10 [0.88, 0.92] - - 1 6
Maximize[ RSL , RSR ] = [ R5 L , R5 R ](1 − [Q1L , Q1R ][Q3 L , Q3 R ])(1 − [Q2 L , Q2 R ][Q4 L , Q4 R ])
subject to
10 exp( x1 2) x2 + 20 x3 + 3 x4 2 + 8 x5 − 200 ≤ 0
(1,1,1,1,1) ≤ ( x1 , x2 , x3 , x4 , x5 ) ≤ (6,3,5, 6, 6)
where
R2 ( x2 ) = 1 − (1 − [0.73, 0.77]) x2
x3 +1
x3 + 1
R3 ( x3 ) = ∑ k
([0.87, 0.89]) ([0.11, 0.13])
x3 +1− k
k =2 k
R4 ( x4 ) = 1 − (1 − [0.68, 0.72]) x4
R5 ( x5 ) = 1 − (1 − [0.83, 0.86]) x5
Maximize[ RSL , RSR ] = [ R5 L , R5 R ](1 − [Q1L , Q1R ][Q3 L , Q3 R ])(1 − [Q2 L , Q2 R ][Q4 L , Q4 R ])
subject to
10 exp( x1 2) x2 + 20 x3 + 3 x4 2 + 8 x5 − 200 ≤ 0
(1,1,1,1,1) ≤ ( x1 , x2 , x3 , x4 , x5 ) ≤ (6, 3, 5, 6, 6)
where
R1 ( x1 ) = {[.8,.8],[.85,.85],[.9,.9],[.925,.925],[.95,.95],[.975,.975]}
R2 ( x2 ) = 1 − (1 − [0.75, 0.75]) x2
x3 +1
x3 + 1
R3 ( x3 ) = ∑ k
([0.88, 0.88]) ([0.12, 0.12])
x3 +1− k
k =2 k
R4 ( x4 ) = 1 − (1 − [0.7, 0.7]) x4
Studies on Reliability Optimization Problems by Genetic Algorithm 68
R5 ( x5 ) = 1 − (1 − [0.85, 0.85]) x5
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
subject to
0.01
C1 ( x) = x1 x2 + 2.2 x2 x3 + 1.5 x2 x4 + 2 exp ≤ 28
1 − R5
0.01
C2 ( x) = x1 + 0.1x2 + 2 x3 + x4 + 5exp ≤ 25
1 − R5
0.01
C3 ( x) = x12 + ( x2 − 2)3 + 1.5 x3 + x4 + 0.6 exp ≤ 21
1 − R5
and Ri = Ri ( xi ) = 1 − (1 − ri ) xi , i = 1, 2, 3, 4, Qi = 1 − Ri , i = 1,..., 5
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
subject to
0.01
C1 ( x) = x1 x2 + 2.2 x2 x3 + 1.5 x2 x4 + 2 exp ≤ 28
1 − R5
Reliability Redundancy Allocation Problems in Interval Environment 69
0.01
C2 ( x) = x1 + 0.1x2 + 2 x3 + x4 + 5exp ≤ 25
1 − R5
0.01
C3 ( x) = x12 + ( x2 − 2)3 + 1.5 x3 + x4 + 0.6 exp ≤ 21
1 − R5
subject to
RS ( x, R) ≥ [0.999, 0.999]
subject to
RS ( x, R) ≥ [0.999, 0.999]
×([Q2 L , Q2 R ] + [ R2 L , R2 R ][Q3 L , Q3 R ]) + [ R3 L , R3 R ][ R5 L , R5 R ][ R6 , R6 ]
subject to
0.01 0.01
C1 ( x) = x1 x2 + 0.5 x1 log(1 + x3 ) + x4 + 2 x5 + 0.3exp + 0.3exp ≤ 27
1 − R6 1 − R7
0.02 0.01
C2 ( x) = ( x1 + 2 x2 + 1.2 x3 ) log(1 + x1 + x2 + 2 x3 ) + 0.4 x4 + 0.2 x5 exp + 0.5exp ≤ 29
1 − R6 1 − R7
r1 = [0.69, 0.71] , r2 = [0.88, 0.92] , r3 = [0.78, 0.81] , r4 = [0.63, 0.66] , r5 = [0.68, 0.71] .
×([Q2 L , Q2 R ] + [ R2 L , R2 R ][Q3 L , Q3 R ]) + [ R3 L , R3 R ][ R5 L , R5 R ][ R6 , R6 ]
subject to
0.01 0.01
C1 ( x) = x1 x2 + 0.5 x1 log(1 + x3 ) + x4 + 2 x5 + 0.3exp + 0.3exp ≤ 27
1 − R6 1 − R7
Studies on Reliability Optimization Problems by Genetic Algorithm 72
0.02 0.01
C2 ( x) = ( x1 + 2 x2 + 1.2 x3 ) log(1 + x1 + x2 + 2 x3 ) + 0.4 x4 + 0.2 x5 exp + 0.5exp ≤ 29
1 − R6 1 − R7
r1 = [0.70, 0.70] , r2 = [0.90, 0.90] , r3 = [0.80, 0.80] , r4 = [0.65, 0.65] , r5 = [0.70, 0.70] .
subject to
RS ( x, R) ≥ [0.999, 0.999]
subject to
RS ( x, R) ≥ [0.999, 0.999]
Reliability Redundancy Allocation Problems in Interval Environment 73
To study the performance of our proposed GAs like PFP-GA and Big-M-GA based on
two different types of penalty techniques, sensitivity analyses (for Example 1) have
been carried out graphically on the centre of the interval valued system reliability
with respect to GA parameters like, population size, crossover and mutation rates
separately keeping the other parameters at their original values. These are shown in
Studies on Reliability Optimization Problems by Genetic Algorithm 74
Figures 3.6, 3.7 and 3.8. From Figure 3.6, it is evident that in case of PFP-GA, a smaller
However, both the GAs are stable when population size exceeds the number
30. From Figure 3.7, it is observed that the system reliability is stable if we consider
the crossover rate between the interval (0.65, 0.95) in case of PFP-GA. In both GAs, it
is stable when crossover rate is greater than 0.8. In Figure 3.8, sensitivity analyses
have been done with respect to the mutation rate. In both GAs, the values of system
1
Centre of the interval valued
objective function
0.75
BIG-M-GA
0.5
PFP-GA
0.25
0
5 10 15 20 25 30 35 40 45 50 55
Population size
Figure 3.6: P_size vs. centre of the objective function for Example 1
0.9
Centre of the interval valued objective
0.896
function
Big-M-GA
0.892
PFP-GA
0.888
0.884
0.6 0.8 1
Cros sover rate
Figure 3.7: P_cross vs. centre of the objective function for Example 1
Reliability Redundancy Allocation Problems in Interval Environment 75
0.5
0 0.05 0.1 0.15 0.2 0.25
Mutation rate
Figure 3.8: P_mute vs. centre of the objective function for Example 1
network system with some resource constraints have been solved. In those systems,
reliability of each component has been considered as an imprecise number and this
representation with known probability distribution, fuzzy set with known fuzzy
membership function or fuzzy number. For handling the resource constraints, the
problem with the help of two different parameter free penalty techniques. Therefore,
with integer and/or mixed-integer variables. To solve the transformed problems, two
different real coded GA-based on different fitness functions have been developed for
mutation for integer variables and boundary mutation for floating point variables)
and elitism of size one. In the existing penalty function technique, tuning of penalty
as these are penalty parameter free techniques. From the performance of GAs, it is
observed that the GAs with both fitness functions due to different penalty techniques
take less CPU times with very small generations to solve the problems. It is clear
from the expression of the system reliability that the system reliability is a
components. Therefore, there is one optimum setup irrespective of the choice of the
upper bound or lower bound of the component reliabilities. As a result, the optimum
setup obtained from the upper bound/lower bound will provide both the upper and
• Introduction
• Assumptions
• Solution Procedure
• Numerical Examples
• Sensitivity Analysis
• Concluding Remarks
Studies on Reliability Optimization Problems by Genetic Algorithm 78
4.1 Introduction
performance specifications. For a system with known cost, reliability, weight, volume
this type is known as the redundancy allocation problem. The basic objective of the
components that either maximizes the system reliability or minimize the system cost
often hard to find feasible solutions. This redundancy allocation problem is NP-hard
and it has been well discussed in Tillman, Hwang and Kuo (1977a) and Kuo and
Prasad (2000). In this chapter, we have formulated two types of redundancy, viz.
component level redundancy known as low-level redundancy and the system level
redundancy known as high-level redundancy for a five-link bridge system where the
To the best of our knowledge, studies of the system reliability with component
reliability as interval valued have already been initiated by Gupta, Bhunia and Roy
environments [Park (1987), Mahapatra and Roy (2006) and Liu (2010)].
Reliability Optimization under High and Low-level Redundancies… 79
type redundancy allocation problem. To find the optimal solution of this type of
problem by GA, order relations of interval numbers assume an important role for GA
operators. Here, we have used our proposed interval order relations discussed in
Chapter 2. Using these we have developed a real coded elitist GA with tournament
proposed problems. Finally, to illustrate the proposed models, for high-level as well
4.2 Assumptions
(iii) The system will not be damaged or failed due to failed components.
(v) The components as well as the system have two different states, viz. operating
(vi) The cost coefficients as well as the amount of resources are imprecise and
interval valued.
components, which give a system design diagram as shown in Figure 4.1, or provide
a total redundant system as shown in Figure 4.2. The component level redundancy is
known as low-level redundancy and the system level redundancy is known as high-
Let us consider a network with n subsystems. The goal of the redundancy allocation
resource constraints and also to minimize the overall system cost subject to the given
Problem 1 Maximize RS ( x)
subject to gi ( x) ≤ bi i = 1, 2, ⋅⋅⋅, m
1 ≤ l j ≤ x j ≤ u j , x j integer, j = 1,..., n
as to maximize the overall system reliability. This problem belongs to the category of
Problem 2 Minimize CS ( x)
subject to RS ( x) ≥ RT
For low-level redundant system (see. Figure 4.1), the corresponding reliability-
subject to gi ( x) ≤ bi i = 1, 2, ⋅⋅⋅, m
1 ≤ l j ≤ x j ≤ u j , x j integer, j = 1,..., n
Reliability Optimization under High and Low-level Redundancies… 81
x
where R j ( x j ) = [ R jL ( x j ), R jR ( x j )] = 1 − (1 − [rjL , rjR ]) j , j = 1, 2, ⋅⋅⋅, n
Problem 4 Minimize CS ( x)
x
where R j ( x j ) = [ R jL ( x j ), R jR ( x j )] = 1 − (1 − [rjL , rjR ]) j , j = 1, 2, ⋅⋅⋅, n
For high-level redundant system (see Figure 4.2), the corresponding reliability-
1 1 1
2 2 2
x1 x2 xn
1 1 1
2 2 2
h h h
Problem 5
Maximize RS ( h) = [ RSL (h), RSR ( h)] = 1 − (1 − f ([r1L , r1R ],...,[ rqL , rqR ],...,[ rnL , rnR ])) h
l ≤ h ≤ u , h integer
Problem 6
Minimize CS (h)
subject to RS (h) ≥ RT
where RS (h) = [ RSL ( h), RSR ( h)] = 1 − (1 − f ([ r1L , r1R ],...,[ rqL , rqR ],...,[ rnL , rnR ])) h
In this section we shall discuss the solution procedure for all the problems
mentioned in earlier section i.e., Problems (3)-(6). These problems are non-linear
function. Using Big- M penalty technique and real coded genetic algorithm with
problems.
subject to gi ( x) ≤ bi , i = 1, 2, ⋅⋅⋅, m
1 ≤ l j ≤ x j ≤ u j , x j integer, j = 1,..., n
[0, 0] if x ∈ S
where θ ( x) =
−[ RSL ( x), RSR ( x )] + [− M , − M ] if x ∉ S
subject to RS ( x) ≥ RT
[0, 0] if x ∈ S
where θ ( x) =
[CSL , CSR ] + [− M , − M ] if x ∉ S
{
and S = x : RS ( x) ≥ RT , and 1 ≤ l j ≤ x j ≤ u j , x j integer, j = 1,..., n }
For the constrained optimization Problem 5
l ≤ h ≤ u , h integer
[0, 0] if h ∈ S
where θ (h) =
−[ RSL (h), RSR (h)] + [− M , − M ] if h ∉ S
subject to RS ( h) ≥ RT
Maximize [Cˆ SL (h), Cˆ SR (h)] = −[CSL (h), CSR (h)] + θ (h) (4.4)
[0, 0] if h ∈ S
where θ (h) =
[CSL , CSR ] + [− M , − M ] if h ∉ S
Here [ Rˆ SL ( x), Rˆ SR ( x)] , [Cˆ sL ( x), Cˆ sR ( x)] , [ Rˆ SL (h), Rˆ SR (h)] and [Cˆ sL (h), Cˆ sR (h)] are the
interval valued auxiliary objective functions. Problems (4.1) and (4.2) are integer
variables x1 , x2 ,..., xn whereas problems (4.3) and (4.4) are integer non-linear
In this section, we have considered the redundancy allocation problem for low-level
redundancy (see Figure 4.3) and for high-level redundancy of five-link bridge system
(see Figure 4.4) for numerical experiments. Bridge system is of use in system
network with subsystem-5 representing a hub and rest of the systems representing
system [Kuo, Prasad, Tillman and Hwang (2001)] works successfully as long as one
fails, then subsystem-5 plays an important role in the system operation. In each
The system reliability of the low-level five-link bridge network system is given by the
expression as follows:
The system reliability of the high-level five-link bridge network system is given by
RS (h) = 1 − (1 − (r1r2 + q2 r3r4 + q1r2 r3r4 + r1q2 q3r4 r5 + q1r2 r3q4 r5 ))h ,
Example 1
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
subject to
5
g1 ( x) = ∑ c j [ x j + exp( x j 4) − b1 ≤ 0
j =1
5
g 2 ( x) = ∑ w j x j exp( x j 4) − b2 ≤ 0
j =1
Studies on Reliability Optimization Problems by Genetic Algorithm 86
Example 2
5 xj
Minimize CS ( x) = ∑ c j x j + exp( )
j =1 4
subject to RS ( x) ≥ RT
where
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
Example 3
Maximize RS (h) = 1 − (1 − (r1r2 + q2 r3r4 + q1r2 r3r4 + r1q2 q3r4 r5 + q1r2 r3q4 r5 )) h
subject to
Reliability Optimization under High and Low-level Redundancies… 87
5
g1 (h) = [ h + exp(h 4)] ∑ c j − b1 ≤ 0
j =1
5
g 2 (h) = [ h exp(h 4)] ∑ w j − b2 ≤ 0
j =1
where ri = [riL , riR ], i = 1, 2,3, 4,5 and qi = [qiL , qiR ] = 1 − [riL , riR ]
Example 4
5
Minimize CS (h) = [ h + exp(h 4)] ∑ c j
j =1
subject to RS (h) ≥ RT
where RS (h) = 1 − (1 − (r1r2 + q2 r3r4 + q1r2 r3r4 + r1q2 q3r4 r5 + q1r2 r3q4 r5 )) h
and ri = [riL , riR ], i = 1, 2,3, 4,5 and qi = [qiL , qiR ] = 1 − [riL , riR ]
All the values of the parameters related to above Examples are given in Table 4.1:
j rj cj b1 wj b2 RT
computational work has been done on a PC with Intel Core-2 duo processor in LINUX
calculate the best found system reliability and best found system cost which are
nothing but the optimal values of system reliability and system cost. Also we have to
Studies on Reliability Optimization Problems by Genetic Algorithm 88
calculated the mean and variance of system reliability as well as system cost. In this
computation, the values of genetic parameters like p_size, max_gen, p_mute and
p_cross have been taken as 100, 100, 0.15 and 0.85 respectively. The computational
It has been observed from the computational results that the mean system
reliability/mean system cost coincides with the best found system reliability/system
cost. This strict coincidence is due to the fact that each trial run provides us the
optimum solution. Also, the lower ends of the standard deviations, measured in
interval form, assume zero value. This speaks of high precision in our optimization
process. It may also be noted that the average CPU time required for implementing
Example 1 3 2 4
x ’s/ h (2,2,1,2,2) (1) (1,1,2,1,1) (3)
Best found
system [0.939545,0.999027] [0.819842,0.928286] - -
reliability
Mean value
[0.939545,0.999027] [0.819842,0.928286] - -
of system
reliability
Mean value
of system - - [53.186,71.904] [102.34,138.159]
cost
Standard
deviation of [0, 0.059482] [0,0.108444] - -
system
reliability
Standard
deviation of - - [0,18.718] [0,35.819]
system cost
analyses have been carried out graphically on the interval valued system reliability
original values. These have been shown in Figure 4.5-Figure 4.8. From Figure 4.5 it
may be observed that both the bounds of the interval valued system reliability are
same for all the values of population size greater than or equal to 30. This implies
that our proposed GA is stable when population size exceeds 30. Similarly, from
Studies on Reliability Optimization Problems by Genetic Algorithm 90
Figure 4.6 it is clear that our proposed GA is stable when maximum number of
generation is greater than or equal to 10. In Figure 4.7 and Figure 4.8, the values of
interval valued system reliability have been computed with respect to the probability
of crossover within the range 0.45 to 0.95 and the probability of mutation within the
range 0.05 to 0.30 respectively. From these figures, it is clear that the proposed GA is
0.96
0.94
0.92
0.9
10 20 30 40 50 60
Population size
Figure 4.5: P_size vs. interval valued system reliability for Example 1
0.96
0.94
0.92
0.9
10 20 30 40 50 60
Max_gen
Figure 4.6: Max_gen vs. interval valued system reliability for Example 1
Reliability Optimization under High and Low-level Redundancies… 91
Low er bound of
0.995 interval valued
reliability
0.965 system reliability
0.95
0.935
0.92
0.45 0.55 0.65 0.75 0.85 0.95
Probability of crossover
Figure 4.7: P_cross vs. interval valued system reliability for Example 1
Low er bound of
0.995 interval valued system
Interval valued system
reliability
0.98 Upper bound of
reliability
0.935
0.92
0 0.05 0.1 0.15 0.2 0.25 0.3
Probability of m utation
Figure 4.8: P_mute vs. interval valued system reliability for Example 1
redundancy and high-level redundancy and proposed four problems where each
interval valued reliability optimization problem. The reduced problem has been
Big-M penalty technique and solved by genetic algorithm. To solve the problem, we
have developed a real coded GA for integer variables with interval valued fitness
mutation and elitism of size one. It is well known that the penalty coefficient plays a
this difficulty, we have used Big-M penalty technique which does not require any
penalty coefficient. This entire approach opens up scope for reliability optimization
when reliability values and other design parameters are interval/imprecise valued.
Thus, it can be claimed that the generalization attempted in this chapter can handle
• Introduction
• Assumptions
• Problem Formulation
• Solution Procedure
• Numerical Example
• Concluding Remarks
Studies on Reliability Optimization Problems by Genetic Algorithm 94
5.1 Introduction
Over the last few decades, attention is being paid to reliability redundancy allocation
problems which have started drawing the attention of the reliability designers for
arriving at reliability optimization designs [Chern (1992), Coit and Smith (1996), Kuo
and Prasad (2000) and Sun and Li (2002)]. The basic objective of a redundancy
a prefixed reliability goal for the system as a whole, subject to several operating
integer programming problem. According to Chern (1992) RAP is NP-hard and it has
been well studied as summarized in Tillman, Hwang and Kuo (1977a) and Kuo and
like heuristic methods [Nakagawa and Nakashima (1977), Kim and Yum (1993) and
Hwang and Kuo (1977b)], reduced gradient method [Hwang, Tillman and Kuo
approach [Kim and Yum (1993)], dynamic programming method [Kuo, Prasad,
Tillman and Hwang (2001)], branch and bound method [Sun and Li (2002)] were
used for solving such RAP. However, these methods have both advantages and
general system as it can be used only for a few particular structures of the objective
function and constraints that are decomposable. In branch and bound method, the
With the advent of genetic algorithm (GA) [Goldberg (1989) and Deb (2000)]
and other evolutionary algorithms, researchers have started paying more attention
Reliability Optimization under Weibull Distribution… 95
to RAP using numerical methods [Coit and Smith (1996, 2002) and Coelho (2009a,
2009b)] as these methods provide more flexibility and require less assumptions on
the objective as well as the constraints and are also effective irrespective of whether
the search space is discrete or not. These have enabled the reliability
the literature in almost all the studies referred above, the design parameters
involved in RAP have usually been taken to be precise values. This means that every
there exist some complete probabilistic information about the system and the
statistical data available in most of the cases where either the system is new or if
exists only as a project. It is not always possible to observe the stability from the
statistical point of view. This means that only some partial information about the
by imprecise probabilities [Coolen and Newby (1994) and Utkin and Gurov (1999,
2001)]. Further, distributional parameters may not be of precise value. They may be
allowed to vary over an interval to take care of the effects of several factors. Keeping
this chapter, we have considered the RAP under imprecise reliability and component
programming problem with interval coefficients for maximizing the overall system
function would be interval valued. For solving such optimization problem by GA, we
arithmetical crossover and boundary mutation for floating point variables. Finally, to
illustrate the proposed model, a numerical example has been solved for different
5.2 Assumptions
(iv) Both the Weibull scale and shape parameters are imprecise and interval valued.
optimize this system reliability under certain constraints, the following topics play
β (t − δ ) β −1 t −δ β
f (t ) = exp − , t ≥δ ≥ 0 .
(θ − δ ) β θ − δ
Now if α = [α L , α R ] and β = [ β L , β R ]
We can easily ensure from interval mathematics that for such a distribution, the
∞
Property-2: ∫ [ f L (t ), f R (t )] dt = [1,1]
−∞
distributed t is given by
n parallel subsystems and mission time for overall system so as to maximize the
overall system reliability subject to the given constraints mostly arriving in linear
Studies on Reliability Optimization Problems by Genetic Algorithm 98
subject to Ax ≤ b
[ βiL , βiR ]
and ri (t ) = [ ri (t ), ri (t )] = e−[α iL ,α iR ]t , i = 1, 2, ⋅⋅⋅, n , under Weibull setup with
In this section we shall discuss the solution procedure for the problem mentioned in
with interval valued objective function. Using Big-M penalty technique this problem
[0, 0] if ( x, t ) ∈ S
where θ ( x, t ) =
−[ RSL ( x, t ), RSR ( x, t )] − [ M , M ] if ( x, t ) ∉ S
five-link bridge network system has been considered (see Figure 5.1).The proposed
method is coded in C programming language and run in the LINUX operating system.
The computational procedure has been implemented on PC with Intel Core-2 duo
processor with 2.5 GHz. For each case, 50 independent runs have been performed to
calculate the best found system reliability which is nothing but the optimal value of
system reliability, mean and standard deviation of system reliability in interval forms
and average CPU time. In this computational work, the values of different genetic
parameters like, population size (p_size), mutation rate (p_mute), crossover rate
(p_cross) and maximum generation (max_gen) have been taken as 200, 0.15, 0.85 and
150 respectively.
Example 1
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
subject to
c( x) = x1 + 0.1x2 + 2.0 x3 + x4 + x5 ≤ 25
w( x) = 2 x1 + x2 + x3 + 0.1x4 + x5 ≤ 21
v( x) = x1 + x2 + x3 + x4 + x5 ≤ 28
where
[ βiL , βiR ]
and ri (t ) = [ riL (t ), riR (t )] = e−[αiL ,α iR ]t , i = 1, 2,3, 4, 5
To study the variation of the parameters, we consider the following four cases
as follows:
Case-I When both the parameters α i and βi are interval valued numbers
Reliability Optimization under Weibull Distribution… 101
Case-II When α i ’s are interval valued and βi ’s are fixed valued numbers
Case-III When α i ’s are fixed valued and βi ’s are interval valued numbers
Case-IV When both the parameters α i and βi are fixed valued numbers
The solutions of five-link bridge network system for different cases have been
displayed in Table 5.2. From Table 5.2 it is seen that the best found values of system
reliability in all cases are the same with mean values of the same. Again in Case-III
and Case-IV the standard deviations of system reliability of the system are zero
whereas in Case-I and Case-II, these are interval valued with lower bounds and
significantly small widths. Also average CPU time required for implementing the
In this chapter, for the first time, the reliability optimization problem with Weibull
Now, for handling the problem with resource constraints, the corresponding problem
has been converted into unconstrained optimization problem with the help of
penalty technique (called Big-M penalty technique [Gupta, Bhunia and Roy (2009)]).
To solve the problem, we have developed a real coded GA for mixed-integer variables
whole arithmetical crossover and boundary mutation for the gene corresponding to
• Introduction
• Assumptions
• Problem Formulation
• Solution Procedure
• Numerical Examples
• Sensitivity Analysis
• Concluding Remarks
Studies on Reliability Optimization Problems by Genetic Algorithm 104
6.1 Introduction
assumed that all the probabilities are precise. This means that every probability
involved is perfectly determinable. In this case it is usually assumed that there exist
determinable.
(ii) The system components are independent, i.e., all the random variables,
During the past, the assumption of uncertainty in most of the methods in reliability is
based on precise probabilities and the reliabilities of the system components are to
be known at a fixed positive number which lies in the open interval (0,1) . The precise
system reliability can be computed theoretically if both the above two conditions are
satisfied (it is assumed that the system structure is defined precisely and there exists
a function linking the system time to failure as well as the times to failure of the
components). If at least one condition is violated, then only the interval measure of
reliability [Barlow and Proschan (1965), Coolen and Newby (1994) and Lindqvist
and Langseth (1998)] can be obtained. However, in real-life situations, there are not
sufficient statistical data in most of the cases where either the system is new or if
exists only as a project. It is not always possible to observe the stability from the
statistical point of view if such data exists. This means that only some partial
interval number [Gupta, Bhunia and Roy (2009)] and is calculated by imprecise
probabilities [Gurov and Utkin (1999) and Utkin and Gurov (1999, 2001)] define
least up to a given period of time under stated conditions. For designing a highly
reliable system, there arises a question as to how to get a balance between the
reliability of a system and resources such as cost, volume and weight. As a result,
In the last two decades, a number of techniques have been proposed for
solving reliability optimization problems [Chern (1992), Gen and Yun (2006) and Ha
method, branch and bound method, Lagrange multiplier method, etc [Kuo, Prasad,
Tillman and Hwang (2001), Sun, Mckinnon and Li (2001) and Sun and Li (2002)]. In
the year 2003, Zhao and Liu (2003) developed stochastic programming technique for
randomness have been discussed with known probability distributions, viz. uniform,
normal and log-normal distributions, when the resource variables are random. The
6.2 Assumptions
(ii) If a component of any subsystem fails to function, the entire system will not be
damaged or failed.
(iv) The state of components and system has only two states like operating state
or failure state.
(v) The resource constraints are chance constraints with resource vector as
stochastic in nature.
parallel-series system or n -stage series system (see Figure 6.1). Assuming all the
n
RS ( x) = [ RSL ( x), RSR ( x )] = ∏ R jL ( x), R jR ( x)
j =1
n
Maximize [ RSL ( x), RSR ( x)] = ∏ R jL ( x), R jR ( x) (6.1)
j =1
subject to
Stochastic Optimization of System Reliability for Series System… 107
Prob[ gi ( x ) ≤ bi ] ≥ 1 − γ i , i = 1, 2,...m
and l j ≤ x j ≤ u j , j = 1, 2,...n .
Stage 1 2 j n
1 1 1 1
2 2 2 2
3 3 3 3
x1 x2 xj xn
1 1 x − µ 2
f ( x; µ , σ ) = exp − ; − ∞ < x , µ < ∞, σ > 0
σ 2π 2 σ
1
; a< x<b
f ( x) = b − a
0 ;otherwise
as X ∼ U ( a, b) .
given by
1 1 log x − µ 2
exp − , x>0
f ( x; µ , σ ) = xσ 2π 2 σ
0, x<0
In this case, bi ∼ U (ξi ,ηi ) . Then the constraint Prob[ gi ( x ) ≤ bi ] ≥ 1 − γ i can be written
ηi
1
where ∫ ηi − ξi dx = 1 − γ i
δi
n
Maximize [ RSL ( x), RSR ( x)] = ∏ R jL ( x), R jR ( x) (6.3)
j =1
subject to
In this case, bi ∼ N ( µbi , σ b2i ) where µbi = mean of bi = E (bi ) and σ b2i = variance of
bi = Var(bi )
where ei is the value of the standard normal variate for which Φ (ei ) = γ i
z
1 x2
Again Φ ( z ) =
2π
∫ exp(−
2
)dx
−∞
n
Maximize [ RSL ( x), RSR ( x)] = ∏ R jL ( x), R jR ( x) (6.4)
j =1
subject to
In this case, bi ∼ LN ( µbi , σ b2i ) where µbi = mean of log(bi ) and σ b2i = variance of log(bi )
where ei is the value of the standard normal variate for which Φ (ei ) = γ i
n
Maximize [ RSL ( x), RSR ( x)] = ∏ R jL ( x), R jR ( x) (6.5)
j =1
subject to
Now we shall solve the deterministic problems (6.3), (6.4) and (6.5) by GA-based
In this section, we shall discuss the solution procedure for all the problems [(6.3),
(6.4) and (6.5)] mentioned in earlier section. These problems are non-linear integer
optimization problems with interval valued objective function. Using Big-M penalty
techniques and real coded genetic algorithm with advanced operators these
[0, 0] if x ∈ S
where θ ( x) =
−[ RSL ( x ), RSR ( x)] + [− M , − M ] if x ∉ S
{ }
and S = x : Prob[ gi ( x) ≤ bi ] ≥ 1 − γ i , i = 1, 2,...m and l j ≤ x j ≤ u j , j = 1, 2,...n be
objective.
To illustrate our proposed GA-based on Big-M penalty technique for solving the
Each example has been formulated using Case-I. In the first example, the reliability of
components are interval valued whereas the second one taken from Yadavalli,
Malada and Charles (2007), the reliabilities of components are fixed. The proposed
computation work has been done on the PC which has Intel Core-2 duo processor
Stochastic Optimization of System Reliability for Series System… 111
with 2.5 GHz. For each example, 20 independent runs have been performed to
calculate the best found system reliability, mean and standard deviation and average
CPU time(s). In this computation, we have taken population size, mutation rate,
crossover rate and maximum generation as 100, 0.15, 0.85 and 50 respectively. The
simulation results have been displayed in Table 6.3. Example 2 has been solved by
our proposed technique expressing the reliability of each component as interval with
Example 1
integer programming problem using the data given in the Table -6.1.
n
Maximize [ RSL ( x), RSR ( x)] = ∏ R jL ( x), R jR ( x)
j =1
subject to
4
Prob[∑ aij x j ≤ bi ] ≥ 1 − γ i , i = 1, 2
j =1
Example 2
integer programming problem using the data given in the Table 6.2.
n
Maximize [ RSL ( x), RSR ( x)] = ∏ R jL ( x), R jR ( x)
j =1
subject to
4
Prob[∑ aij x j ≤ bi ] ≥ 1 − γ i , i = 1, 2
j =1
sensitivity analyses have been done graphically on the interval valued system
Stochastic Optimization of System Reliability for Series System… 113
their original values. These are shown in Figure 6.2-Figure 6.5. The graphs (Ref.
Figure 6.2-Figure 6.5) have been drawn for lower and upper bounds of the system
reliability in the same graph. In Figure 6.2, the effect of population size (p_size) on the
system reliability has been studied from the range 10 to 90 of population size
(p_size). In this study, it is observed that both the bounds of the interval be the same
for all values of p_size after 30. This means that our proposed GA is stable when
population size exceeds the number 30. In Figures 6.3-6.5, the values of system
within the range from 0.35 to 0.95, probability of mutation (p_mute) within the range
0.05 to 0.3 and maximum number of generation (max_gen) respectively. From these
generation (max_gen).
1.05
1
System Reliability
0.8
10 20 30 40 50 60 70 80 90
P_size
Figure 6.2: P_size vs. interval valued system reliability for series system
Studies on Reliability Optimization Problems by Genetic Algorithm 114
1.05
System Reliability
1 Lower bound of the
interval valued
0.95 system reliability
0.9 Upper bound of the
interval valued
0.85 system reliability
0.8
0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
P_cross
Figure 6.3: P_cross vs. interval valued system reliability for series system
1.05
1
System Reliability
Lower bound of
0.95 the interval valued
s ys tem reliability
0.9 Upper bound of
the interval valued
0.85 s ys tem reliability
0.8
0 0.05 0.1 0.15 0.2 0.25 0.3
P_mute
Figure 6.4: P_mute vs. interval valued system reliability for series system
1.05
1
System Reliability
0.8
10 20 30 40 50 60 70
Max_gen
Figure 6.5: Max_gen vs. interval valued system reliability for series system
Stochastic Optimization of System Reliability for Series System… 115
system with some resource constraints have been formulated and solved considering
variable representation with known probability distribution, fuzzy set with known
fuzzy membership function or fuzzy number. For handling the resource constraints,
problem with the help of our developed penalty technique (called Big-M penalty
have developed a real coded GA for integer variables with interval valued fitness
and elitism of size one. In tournament selection process and elitism operation we
have used the definitions of interval order relations. It is well known that the penalty
difficulty, we have also proposed Big-M penalty technique which does not require
any penalty coefficient. Here, we have also proposed a new mutation scheme (called
one-neighborhood mutation) in which the selected gene will take either its next
number or its previous number (if exists). In the study of statistical analysis for
interval valued numbers, it has been observed that for the set of same interval valued
numbers, the standard deviation will be an interval with negligible width. However,
• Introduction
• Assumptions
Chance Constraints
• Solution Procedure
• Numerical Example
• Concluding Remarks
Reliability Optimization with Interval Parametric Values… 117
7.1 Introduction
In almost all the studies referred in earlier Chapters 3-6, the design parameters
involved in the optimization problem have usually been taken to be constants. In that
mostly, the design parameters are not deterministic in nature. While in-house
from the factor market, especially in respect of cost can hardly be deterministic.
These parameters can be viewed as estimated values, which in turn follow certain
optimization problems are usually solved in the deterministic domain and need to be
solved in the stochastic domain. For results in the stochastic domain, one may refer
to the works of Coit, Jin and Wattanapongsakorn (2004), Zafiropoulos and Dialynas
(2004) and Marseguerra, Zio, Podofillini and Coit (2005). Further, the distributional
parameters may not be of single value. They may be allowed to vary over an interval
parameters assuming interval values. The resultant solutions will have a greater
appeal, so far as their applications are concerned, because interval values will make
imprecise and/or interval valued have already been initiated by some authors
[Coolen and Newby (1994), Utkin and Gurov (1999, 2001) and Gupta, Bhunia and
the optimum hot redundancy allocation problem under imprecise reliability with
Studies on Reliability Optimization Problems by Genetic Algorithm 118
random variables with interval valued parameters. In fact, even for the random
with interval valued means and variances so that the optimization problem can be
dealt with under a generalized setup. Specific solutions obtained in earlier studies
7.2 Assumptions
(i) The reliability of some of the components are imprecise and interval valued.
(ii) The chance of failure of any component is statistically independent with respect
(iv) All the resource coefficients and available resource amounts are random in
nature.
(v) Mean and variance of each random variable are interval valued.
function is given by
1 1 x − µ 2
f ( x; µ , σ ) = exp − ; − ∞ < x < ∞, − ∞ < µ < ∞ , σ > 0
σ 2π 2 σ
X −µ
If X ∼ N ( µ , σ 2 ) , then Z = , is called a standard normal variate with mean zero
σ
Φ (− z ) = 1 − Φ ( z ), z > 0
b−µ a−µ
Prob(a ≤ X ≤ b) = Φ ( ) − Φ( ) , where X ∼ N ( µ , σ 2 )
σ σ
But mostly these parameters are estimated from a given set of observations. In case
of confidence intervals proposed for the parametric measures one has to examine the
mind we have introduced interval valued normal distribution with both the mean
can easily ensure, from interval arithmetic, that for such a distribution, the following
property holds.
b − µL a − µU
Property-1: Prob(a ≤ X ≤ b) = Φ ( ) − Φ( )
σL σU
b − [ µ L , µU ] a − [ µ L , µU ]
Prob( a ≤ X ≤ b) = Φ ( ) − Φ( )
[σ L , σ U ] [σ L , σ U ]
b − µU b − µ L a − µU a − µ L
= Φ ([ , ]) − Φ ([ , ])
σU σL σU σL
Now, from the monotonically increasing nature of Φ (.) function and the standard
b − µL a − µU
Prob(a ≤ X ≤ b) = Φ ( ) − Φ( )
σL σU
This property of interval valued normal distribution will be of use for deterministic
Chance Constraints
Let us introduce the problem of reliability optimization under chance constraints. Let
reliability levels of the rest ( n − q ) general subsystems which are of precise reliability
values so as to maximize the overall system reliability subject to the given chance
subject to {
}
Prob C j′ g j ( x, R) ≤ b j ≥ 1 − γ j , j = 1, 2,..., m (7.1)
c j1 g j1 ( x, R)
c j2 g j 2 ( x, R )
. .
where C j = , g j ( x, R ) =
. .
. .
c jn g jn ( x, R)
j j
Reliability Optimization with Interval Parametric Values… 121
Since each g j (.) is a technical relationship no such distribution has been assumed for
the same.
nj
Let us denote h j = ∑ c jk g jk ( x, R) − b j , j = 1, 2,..., m .
k =1
{ }
Prob h j ≤ 0 ≥ 1 − γ j , j = 1, 2,..., m (7.2)
random variables c jk ’s and b j then each h j ( j = 1, 2,..., m) will also follows normal
distribution.
nj
E ( h j ) = ∑ E (c jk ) g jk − E (b j ) , j = 1, 2,..., m
k =1
given by
nj
Var(h j ) = ∑ g jk 2 var(c jk ) + var(b j ) , j = 1, 2,..., m (7.3)
k =1
{ }
Then the constraint Prob h j ≤ 0 ≥ 1 − γ j can be written as
− E (h j )
Φ( ) ≥ Φ (e j ) (7.4)
var(h j )
E ( h j ) + e j var(h j ) ≤ 0 , j = 1, 2,..., m .
nj nj
or, ∑ E (c jk ) g jk + e j ∑ g jl′2 var(c jp ) + var(b jp ) ≤ E (b j ) , j = 1, 2,..., m (7.5)
k =1 p =1
nj nj
nj nj nj nj
or, ∑ c jkL g jk + e j ∑ g jp 2c jpL + b jL , ∑ c jkR g jk + e j ∑ jp jpR jR ≤ [b jL , b jR ] ,
g 2
c + b
k =1 p =1 k =1 p =1
j = 1, 2,..., m (7.6)
subject to
nj nj nj nj
∑ c jkL g jk + e j ∑ g jp c jpL + b jL , ∑ c jkR g jk + e j
2
∑ jp jpR jR ≤ [b jL , b jR ] ,
g 2
c + b
k =1 p =1 k =1 p =1
j = 1, 2,..., m .
1 ≤ li ≤ xi ≤ ui , xi integer, i = 1,..., q,
bridge network system as given in Figure 7.1. Bridge system is of use in system
network with subsystem-5 representing a hub and rest of the systems representing
This five-link bridge network system [Kuo, Prasad, Tillman and Hwang (2001)]
(2, 3) fails, then subsystem-5 plays an important role in the system operation. As a
result, estimation of its parameters is to be made with the highest precision and this
of subsystem-5.
The system reliability of the bridge network system is given by the expression as
follows:
where Ri = 1 − Qi , i = 1, 2, 3, 4
In this section we shall discuss the solution procedure for the problem mentioned in
with interval valued objective. Using Big-M penalty technique this problem is
[0, 0] if ( x, R ) ∈ S
where θ ( x, R) =
−[ RSL ( x, R), RSR ( x, R)] + [− M , − M ] if ( x, R) ∉ S
{ { }
and S = ( x, R) : Prob C j′ g j ( x, R) ≤ b j ≥ 1 − γ j , j = 1, 2,..., m and l ≤ x ≤ u , L ≤ R ≤ U
}
where l = (l1 , l2 ,..., lq ) , u = (u1 , u2 ,..., uq ) , x = ( x1 , x2 ,..., xq ) , L = ( Lq +1 , Lq + 2 ,..., Ln ),
This problem can easily be solved by real coded genetic algorithm and interval order
relations.
five-link bridge network system under chance constraints has been considered for
language and run in a LINUX environment. The computational work has been done
on a PC with Intel Core-2 duo processor and 2.5 GHz. For each case, 20 independent
runs have been performed to calculate the best found system reliability, mean,
standard deviation of system reliability and average CPU time for different sets of
population size, mutation rate, crossover rate and maximum number of generations
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
subject to
0.01
Prob(c11 x1 x2 + c12 x2 x3 + c13 x2 x4 + 2 exp ≤ b1 ) ≥ 1 − γ 1
1 − R5
0.01
Prob(c21 x1 + c22 x2 + c23 x3 + c24 x4 + 5exp ≤ b2 ) ≥ 1 − γ 2
1 − R5
Studies on Reliability Optimization Problems by Genetic Algorithm 126
0.01
Prob(c31 x12 + c32 ( x2 − 2)3 + c33 x3 + c34 x4 + 0.6 exp ≤ b3 ) ≥ 1 − γ 3
1 − R5
c11 ∼ N ([1,1],[0, 0]) , c12 ∼ N ([2.2, 2.2],[0, 0]) , c13 ∼ N ([1.5,1.5],[0, 0]) ,
b1 ∼ N ([28, 28],[0, 0]) , c21 ∼ N ([1,1],[0, 0]) , c22 ∼ N ([0.1, 0.1],[0, 0]) ,
c23 ∼ N ([2, 2],[0, 0]) , c24 ∼ N ([1,1],[0, 0]) , b2 ∼ N ([25, 25],[0, 0]) ,
(ii) All the random variables c jk ( j = 1, 2, 3; k = 1, 2,3, 4) and b j ( j = 1, 2,3) have interval
c11 ∼ N ([0.9,1.1],[0, 0]) , c12 ∼ N ([2.1, 2.4],[0, 0]) , c13 ∼ N ([1.4,1.6],[0, 0]) ,
b1 ∼ N ([26, 29],[0, 0]) , c21 ∼ N ([0.8,1.1],[0, 0]) , c22 ∼ N ([0.09, 0.11],[0, 0]) ,
c23 ∼ N ([1.9, 2.1],[0, 0]) , c24 ∼ N ([0.8,1.2],[0, 0]) , b2 ∼ N ([23, 27],[0, 0]) ,
(iii) All the random variables c jk ( j = 1, 2,3; k = 1, 2,3, 4) and b j ( j = 1, 2, 3) have interval
c11 ∼ N ([0.9,1.1],[0.1, 0.1]) , c12 ∼ N ([2.1, 2.4],[0.1, 0.1]) , c13 ∼ N ([1.4,1.6],[0.1, 0.1]) ,
γ 2 = 0.05
γ 3 = 0.05
(iv) All the random variables c jk ( j = 1, 2, 3; k = 1, 2,3, 4) and b j (i = 1, 2,3) have fixed
c11 ∼ N ([1,1],[0.1, 0.2]) , c12 ∼ N ([2.2, 2.2],[0.1, 0.15]) , c13 ∼ N ([1.5,1.5],[0.1, 0.3]) ,
c21 ∼ N ([1,1],[0.1, 0.2]) , c22 ∼ N ([0.1, 0.1],[0.01, 0.02]) , c23 ∼ N ([2, 2],[0.1, 0.3]) ,
(v) All the random variables c jk ( j = 1, 2, 3; k = 1, 2,3, 4) and b j ( j = 1, 2,3) have interval
γ 3 = 0.05
Optimum solution sets, as obtained via genetic algorithm for the six possible
optimization problem keeping the same parametric values of Sun, Mckinnon and Li
(2001). Similar study under the proposed setup is carried out in Table-7.2.
+[Q1L , Q1R ][ R2 L , R2 R ][ R3 L , R3 R ][ R4 L , R4 R ]
subject to
0.01
x1 x2 + 2.2 x2 x3 + 1.5 x2 x4 + 2 exp ≤ 28
1 − R5
0.01
x1 + 0.1x2 + 2 x3 + x4 + 5exp ≤ 25
1 − R5
0.01
x12 + ( x2 − 2)3 + 1.5 x3 + x4 + 0.6 exp ≤ 21
1 − R5
If we take r1 = [0.70, 0.70] , r2 = [0.85, 0.85] , r3 = [0.75, 0.75] , r4 = [0.8, 0.8] then
Table-7.2 presents comparative results obtained from the proposed method and
( x, R5 ) RS CPU seconds
From the above table, it is observed that the system reliability has been increased
and at the same time, the processing time has also been decreased.
For the first time, we have examined the reliability optimization problem in the
stochastic domain with respect to available resources and in the interval domain
with respect to the component reliabilities and employed the genetic algorithm to
arrive at an optimum solution set. It has been observed from the computational
results, under taken herein for a five-link bridge network, that the mean system
reliability coincides with the best found system reliability. This strict coincidence is
due to the fact that each trial run provides us optimum solution in our study. Also,
the lower ends of the standard deviations, measured in interval form, assume zero
value. It may also be noted that the average CPU time required for implementing the
genetic algorithm, is also on the lower side. Further, optimization under stochastic
• Introduction
• Assumptions
• Tchebycheff Problem
• Lexicographic Ordering
• Lexicographic Problem
• Problem Formulation
• Solution Procedure
• Numerical Example
• Sensitivity Analysis
• Concluding Remarks
Studies on Reliability Optimization Problems by Genetic Algorithm 132
8.1 Introduction
function. Mostly the researchers have framed the reliability optimization problem as
formulation to maximize the system reliability and minimize the cost for reliability
overview of the trend of research in this area, one may refer to the works of Park
(1987), Dhingra (1992), Rao and Dhingra (1992), Srinivas and Deb (1994), Ravi,
Reddy and Zimmermann (2000), Huang, Tian and Zuo (2005), Coit and Konak (2006)
and others. In the recent years, Taboada and Coit (2007) proposed a new method
Pareto optimal set for multi-objective reliability optimization design problems. In the
first approach, they developed pseudo-ranking scheme to select the solutions by the
they used data mining clustering techniques to group the data by using k-means
algorithm to find the clusters of similar solutions. Ramirez-Marquez and Coit (2007a)
in multi-state systems. In the same area, Taboada, Espiritu and Coit (2008a)
Multi-objective Reliability Optimization in Interval Environment 133
series-parallel system for power system. Taboada, Espiritu and Coit (2008b) solved
maximizing the system reliability and minimizing both the system cost and weight.
In the year 2009, Li, Liao and Coit (2009) proposed a two-stage approach for multi-
In the single objective optimization one attempts to obtain the best design or
hand for the multiple objectives, there may not exist one solution which is best
optimization, there exists a set of solutions which are superior to the rest of the
solutions in the search space when all the objectives are considered, but are inferior
to other solutions in the space in one or more objectives. These solutions are known
and the rest of the solutions are known as dominated solutions. Since none of the
solutions in the non-dominated set is absolutely better than any other, any one of
Studies on Reliability Optimization Problems by Genetic Algorithm 134
system reliability would be interval valued. In this chapter, we have proposed GA-
with interval objectives. The objectives considered are the maximization of the
system reliability and minimization of the system cost. Here also, we have considered
the interval valued cost coefficients. For this purpose we have formulated several
valued objectives. In this connection, we have also proposed the definition of Pareto
function technique. For solving such typical problems, we have developed a real
solved and to investigate the overall performance of the proposed GA-based penalty
8.2 Assumptions
(iii) The system will not be damaged or failed due to failed components.
(v) The components as well as the system have two different states, viz. operating
According to the existing literature, there are several methods developed for solving
now, none has developed the techniques/ methods for solving multi-objective
optimization problem with interval valued objectives for several decision variables.
subject to x ∈ S
and S = {x : g j ( x) ≤ 0, j = 1, 2, ⋅⋅⋅, m}
decision-makers’ point of view) and ideal objectives and different types of ideal
Definition: A decision vector x* ∈ S is Pareto optimal if there does not exist another
Definition: Let X be a metric space. The (open) ball of radius δ > 0 centered at a
function or metric. If the less than symbol (<) is replaced by a less than or equal
ball: B( x* , δ ) = {x ∈ X : d ( x, x* ) ≤ δ } .
Definition: A decision vector x* ∈ S is locally Pareto optimal if there exists δ > 0 such
centered at a point x* .
Definition: A decision vector x* ∈ S is weakly Pareto optimal if there does not exist
whose components are formed by zi** = zi* − ε i for all i = 1, 2, ⋅⋅⋅, k , where zi* is the
component of the ideal objective vector and ε i > 0 is a relatively small but
Definition: Let X = » n and suppose that ξ = {ξ1 , ξ 2 , ⋅⋅⋅, ξ n } and η = {η1 ,η2 , ⋅⋅⋅,η n } be
follows:
Multi-objective Reliability Optimization in Interval Environment 137
1
n pp
d p (ξ ,η ) = ∑ ξi − ηi where 1 ≤ p < ∞
i =1
and d ∞ (ξ ,η ) = max { ξi − ηi }
1≤i ≤ n
Definition: Let X = l p , 1 ≤ p < ∞ , be the set of all sequences ξ = {ξi } of real scalars
1
∞ n pp
∑ ξi < ∞ . Define the mapping d : X × X → » by d (ξ ,η ) = ∑ ξi − ηi
p
such that
i =1 i =1
According to the existing literature there are several techniques for solving the multi-
Now, we shall formulate all these problems with interval valued objectives.
Maximize Ai ( x)
Studies on Reliability Optimization Problems by Genetic Algorithm 138
subject to x ∈ S
and obtain the optimal value ( say,) zi* = [ ziL* , ziR* ] for i = 1, 2, ⋅⋅⋅, k .
Step 2: Using the above reference point and d p -metric used for measuring, we get
1
k p p
Minimize ∑ [ fiL ( x) − ziR* , fiR ( x) − ziL* ] (8.1)
i =1
subject to x ∈ S
1 1
The exponent may be dropped. Problems with or without the exponent are
p p
Tchebycheff Problem
form:
Minimize Max
i =1,2,⋅⋅⋅,k
( A ( x) − z )
i
*
i (8.2)
subject to x ∈ S
(
Minimize Max wi Ai ( x) − zi*
i =1,2,⋅⋅⋅,k
) (8.3)
k
subject to ∑ wi = 1
i =1
and x ∈ S
Lexicographic Ordering
to their absolute importance. This means that the more important objective is
infinitely more important. After ordering, the most important objective function is
optimized subject to the given constraints. If this problem has a unique solution, it is
Lexicographic Problem
Let the objective functions be sorted according to the lexicographic order from the
most important to the less important. In this technique, the given multi-objective
subject to x ∈ S
k
i =1,2,⋅⋅⋅,k
( i =1
) (
lex Minimize Max wi Ai ( x) − zi* , ∑ Ai ( x) − zi**
) (8.5)
k
subject to ∑ wi = 1
i =1
and x ∈ S
where zi** = zi* − ε i is the i-th component of utopian objective vector which is an
reliability of each component as well as the cost of resources are interval valued.
system, the system reliability and also the system cost would be interval valued.
Assuming all the components in i-th subsystem as identical, the system reliability RS
is given by
n
RS ( x) = [ RSL ( x), RSR ( x) ] = ∏ [ RiL ( x), RiR ( x) ]
i =1
i = 1, 2,..., n by maximizing the system reliability [ RSL ( x), RSR ( x)] and minimizing the
system cost [CSL ( x), CSR ( x)] , subject to the given constraints. Hence the problem can
be written as
Multi-objective Reliability Optimization in Interval Environment 141
n
Maximize [ RSL ( x), RSR ( x )] = ∏ [ RiL ( x), RiR ( x)] (8.6)
i =1
Stage 1 2 i n
1 1 1 1
2 2 2 2
3 3 3 3
x1 x2 xi xn
objectives.
As the objective functions of (8.6) are interval valued, we have to modify the existing
methods for solving the said problem. In global criteria method for solving MOOPs
with fixed (non-interval) objective, the ideal objective vector is used as a reference
individually subject to the constraints. In problem (8.6), the objective functions are
interval valued. Hence, in this case, each component of ideal objective vector would
be interval valued. Hence, different steps of modified global criteria method are as
follows:
Studies on Reliability Optimization Problems by Genetic Algorithm 142
subject to g k ( x) ≤ 0, k = 1, 2,..., m
Step 2:
subject to g k ( x) ≤ 0, k = 1, 2,..., m
* *
and find the optimal value ( say,) [CSL , CSR ].
* * * *
Step 3: Form the ideal objective vector ([ RSL , RSR ],[CSL , CSR ]) with the interval valued
components.
Step 4: Using the above reference point and d p -metric used for measuring.
Minimize [− RSR ( x) − RSR* , − RSL ( x) − RSL* ] + [CSL ( x) − CSR* , CSR ( x) − CSL* ] (8.7)
p p p
subject to g k ( x) ≤ 0, k = 1, 2,..., m
1 1
The exponent may be dropped. Problems with or without the exponent are
p p
form:
(
Minimize Max [− RSR ( x) − RSR* , − RSL ( x) − RSL* ] , [CSL ( x) − CSR* , CSR ( x) − CSL* ] ) (8.8)
Multi-objective Reliability Optimization in Interval Environment 143
subject to g j ( x) ≤ 0, j = 1, 2,..., m
Minimize
(
Max w [− RSR ( x) − RSR* , − RSL ( x) − RSL* ] , (1 − w) [CSL ( x) − CSR* , CSR ( x) − CSL* ] ) (8.9)
subject to g j ( x) ≤ 0, j = 1, 2,..., m
Let the objective functions be arranged according to the lexicographic order from the
most important [− RSR ( x), − RSL ( x)] to the less important[CSL ( x), CSR ( x)] . In this
lex Minimize ([− RSR ( x), − RSL ( x)] ,[CSL ( x), CSR ( x)] ) (8.10)
subject to g j ( x) ≤ 0, j = 1, 2,..., m
(
lex Minimize Max w [− RSR ( x) − RSR* , − RSL ( x) − RSL* ] , (1 − w) [CSL ( x) − CSR* , CSR ( x) − CSL* ] , )
([ − R
SR ( x ) − RSR
**
, − RSL ( x) − RSL** ] + [CSL ( x ) − CSR** , CSR ( x) − CSL** ] )
subject to g j ( x) ≤ 0, j = 1, 2,..., m . (8.11)
** ** ** **
where ([ RSL , RSR ],[CSL , CSR ]) is the utopian objective vector which is an infeasible
* * * *
([ RSL − ε1R , RSR − ε1L ], [CSL − ε 2 R , CSR − ε 2 L ]) , where [ε1L , ε1R ], [ε 2 L , ε 2 R ] are
In this section we shall discuss the solution procedure for all the problems
problem with interval valued objectives. Using Big-M penalty technique these
any given constrained optimization problem with an interval valued fitness function
penalizing a large positive number say, M which can be written in the interval form
g j ( x) ≤ 0, j = 1, 2,..., m
[0, 0] if x ∈ S
where θ ( x) =
[ f auxL , f auxR ] − [ M , M ] if x ∉ S
{ }
and S = x : g j ( x) ≤ 0, j = 1, 2,..., m be the feasible space.
Here ( −[ f auxL , f auxR ]) is the interval valued auxiliary objective function. Problem
Example 1
5
Maximize [ RSL ( x), RSR ( x)] = ∏ 1 − [1 − riR ,1 − riL ]xi
i =1
5
Minimize [CSL ( x), CSR ( x)] = ∑ [CiL , CiR ][ xi + exp( xi 4)]
i =1
5
g1 ( x) = ∑ Pi xi 2 − b1 ≤ 0
i =1
5
g 2 ( x) = ∑ Wi [xi exp( xi 4)] − b2 ≤ 0
i =1
b1 and b2 are given in Table 8.1. The proposed method/technique has been coded in
C programming language. The computational work has been done on a PC with Intel
Core-2-duo 2.5 GHz processor in LINUX environment. For each case 20 independent
runs have been performed to calculate the best found system reliability which is
nothing but the optimal value of the system reliability. In this computation, the
values of genetic parameters like, population size (p_size), mutation rate (p_mute),
crossover rate (p_cross) and maximum number of generations (max_gen) have been
taken as 100, 0.15, 0.85 and 150 respectively. The computational results have been
i 1 2 3 4 5
ri [0.78, 0.82] [0.84, 0.85] [0.87, 0.91] [0.63, 0.66] [0.74, 0.76]
Pi 1 2 3 4 2
Wi 7 8 8 6 9
b1 = 110 , b2 = 200
(ii) the best found value, mean value of system reliability R* and the corresponding
system cost C * in Lexicographic problem are higher than the same in other
problems.
(iii) all the results in Tchebycheff problems are greater than weighted Tchebycheff
Hence from the above observation, it can be concluded that the solution of
Lexicographic problem is the best solution. In this case, the best found value as well
as mean value of system reliability R* are far away from the same results obtained
system cost, in that case, he/she may take the solution of either weighted
same solution.
Multi-objective Reliability Optimization in Interval Environment 147
problems, sensitivity analyses have been carried out graphically on the system
parameters at their original values. These have been shown in Figures 8.2- 8.4. From
Figure 8.2, it is observed that both the bounds of the system reliability be the same
for all the values of population size (p_size) greater than or equal to 60. This means
that our proposed GA is stable when population size exceeds 60. In Figures 8.3 and
8.4, the values of system reliability have been computed with respect to the
probability of crossover (p_cross) within the range from 0.55 to 0.95 and the
probability of mutation (p_mute) within the range 0.05 to 0.25 respectively. From
these figures, it is evident that the proposed GA is stable with respect to probability
System Reliability
Upper limit of interval
0.92 valued system reliability
0.88
0.84
0.8
50 60 70 80 90 100 110 120
Popsize
Figure 8.2: P_size vs. interval valued system reliability for MOOP
1 Lower limit of
interval valued
system reliability
0.9
System reliability
Upper limit of
interval valued
0.8 system reliability
0.7
0.6
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Probability of Crossover
Figure 8.3: P_cross vs. interval valued system reliability for MOOP
1
Lower limit of interval
valued system reliability
0.9
System Reliability
0.6
0.5
0.05 0.10 0.15 0.20 0.25
Probability of Mutation
Figure 8.4: P_mute vs. interval valued system reliability for MOOP
Multi-objective Reliability Optimization in Interval Environment 149
environment. For this purpose, we have proposed the definition of Pareto optimality
minimizing system cost under the assumption that the reliability of each component
as well as the cost coefficients are interval valued. Then the problem has been
reduced problem has been converted into single objective optimization problem
using Big-M penalty technique and solved by genetic algorithm and our newly
• General Conclusion
In this thesis, for the first time, we have investigated different types of reliability
optimization problems in interval environment under the assumption that either the
system and reliability network system with some resource constraints have been
solved. We have also formulated and solved two different redundancies known as
5, the reliability optimization problem with Weibull distributed (with interval valued
resource constraints have been solved. On the other hand chance constrained
have been formulated and solved in Chapter 6. We have also examined the reliability
interval objective functions. In the whole work of the thesis, the reliability of each
function is converted into interval valued objective. To solve those problems, the
interval order relations are very essential. For this purpose, we have proposed a new
Studies on Reliability Optimization Problems by Genetic Algorithm 152
Chanas and Kuchata (1996), Kundu (1997), Sengupta and Pal (2000) and Mahato and
Bhunia (2006).
For further researches a lot of scope may arise from this thesis.
(i) The proposed techniques may be applied for solving real-life decision-making
(ii) In this thesis, we have solved all the optimization problems with the help of
genetic algorithm with interval valued fitness function. These problems can be
(iii) In this thesis, we have used Big-M penalty and PFP penalty techniques to solve
(iv) The proposed approach of Chapter 5 opens up the scope for reliability
interval sets.
General Conclusions and Scope of Future Research 153
problems considering only two objectives, viz. system reliability and cost. One
may extend the problem for higher objectives, viz. system reliability, cost,
volume and weight. The same methodologies can be applied to solve the multi-
different types of fitness function. In this connection, one may develop other
Cliffs, NJ.
Allella, F., Chiodo, E. and Lauria, D. (2005), Optimal reliability allocation under
Wiley.
Barlow, R. E. and Proschan, F. (1975), Statistical Theory of Reliability and Life Testing:
Bellman, R. and Dreyfus, S. (1958), Dynamic programming and the reliability of multi
University Press.
614-619.
Chen, T. (2006), IAs based approach for reliability redundancy allocation problems,
510.
Coit, D. W. and Konak, A. (2006), Multiple weighted objectives heuristic for the
Davis, L. (1991), Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York.
582.
Fyffee, D. E., Hines, W. W. and Lee, N. K. (1968), System reliability allocation and a
Gen, M. and Cheng, R. (1997), Genetic Algorithms and Engineering Design, New York:
Gen, M. and Cheng, R. (2000), Genetic algorithms and engineering optimization, John
Gen, M. and Yun, Y. (2006), Soft computing approach for reliability optimization:
Gen, M., Ida, K., Tsujimura, Y. and Kim, C. E. (1993), Large-scale 0-1 fuzzy goal
Gopal, K., Aggarwal, K. K. and Gupta, J. S. (1980), A New Method for solving
Ha, C. and Kuo, W. (2006a), Multi-Path heuristic for redundancy allocation: the tree
480.
Ida, K., Gen, M. and Yokota, T. (1994), System reliability optimization with several
267.
219-225.
Publishers, Netherlands.
12, 354-363.
Kim, J. and Yum, B. (1993), A heuristic method for solving redundancy optimization
Kuo, W. and Wan, R. (2007b), Recent advances in optimal reliability allocation, IEEE
Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 37,
143-156.
Kuo, W., Lin, H., Xu, Z. and Zhang, W. (1987), Reliability Optimization with the
Kuo, W., Prasad, V. R., Tillman, F. A. and Hwang, C. L. (2001), Optimal Reliability
380-388.
Li, C. Y., Chen, X., Yi, X. and Tao, J. (2010), Heterogeneous redundancy optimization
202-207.
Li, Z., Liao, H. and Coit, D. W. (2009), A two-stage approach for multi-objective
1592.
Liang, Y. and Smith, A. E. (2004), An ant colony optimization algorithm for the
reliability design problems using feature models-A concept for early design
Lirong C., Kuo, W., Loh, H. T. and Xie M. (2004), Optimal allocation of minimal &
193-199.
Marseguerra, M., Zio, E., Podofillini, L. and Coit, D. W. (2005), Optimal design of
Mettas, A. (2000), Reliability allocation and optimization for complex systems, IEEE
Meziane, R., Massim, Y., Zeblah, A., Ghoraf, A. and Rahil, R. (2005), Reliability
Publishers.
Nahas, N., Nourelfath, M. and Ait-Kadi, D. (2007), Computing ant colony and the
30, 175-180.
Nakagawa, Y., Nakashima, K., and Hattori, Y. (1978), Optimal Reliability Allocation by
Nourelfath, M., and Nahas, N. (2005), Artificial neural networks for reliability
Ohtagaki, H., Nakagawa, Y., Iwasaki, A. and Narihisa, H. (1995), Smart greedy
Onishi, J., Kimura, S., James, R. J. W and Nakagawa, Y. (2007), Solving the Redundancy
based Methods for the Design and Operation of Technological Systems and
Processes, 10-20.
Ravi, V., Reddy, P. J., and Zimmermann, H. J. (2000), Fuzzy global optimization of
models, IEEE Transactions on System, Man, and Cybernetics, Part A, 28, 691-698.
series parallel system by the surrogate worth trade off method, Microelectronics
173-174.
Academic Publishers.
Sasaki, M. and Gen, M. (2003a), Fuzzy multiple objective optimal system design by
Sun, X. and Li Duan, D. (2002), Optimal Condition and Branch and Bound Algorithm
Engineering, 3, 53-65.
Sun, X. L., Mckinnon, K. I. M. and Li, D. (2001), A convexification method for a Class of
Sung, C. S. and Cho, Y. K. (1999), Branch and Bound redundancy optimization for a
48, 108-117.
Taboada, H., Baheranwala, F., Coit, D.W. and Wattanapongsakorn, N. (2007), Practical
Tian, Z. and Zuo, M. J. (2006), Redundancy allocation for multi-state systems using
Tillman, F. A., Hwang, C. L. and Kuo, W. (1977a), Optimization technique for system
155.
Technology, 110-139.
Vose, M. D. (1999), The Simple Genetic Algorithm: Foundations and Theory, MIT Press,
Cambridge, MA.
Xu, Z., Kuo, W. and Lin, H. (1990), Optimization Limits in Improving System
Yalaoui, A., Chatelet, E. and Chu, C. (2005), A new dynamic programming method for
You, P. and Chen, T. (2005), An efficient heuristic for series parallel redundant
Yu, H., Yalaoui, F., Chatelet, E. and Chu, C. (2007), Optimal design of a maintainable
Yun, W. Y., Song, Y. M. and Kim, H. (2007), Multiple multi-level redundancy allocation
Zhao, J., Liu, Z. and Dao, M. (2007), Reliability optimization using multiobjective ant
colony system approaches, Reliability Engineering and System Safety, 92, 109-
120.
Zhao, R. and Liu, B. (2003), Stochastic programming models for general redundancy-
1476-1481.
Zhao, Xian, Cui, L and Kuo, W. (2007), Reliability for Sparsely Connected Consecutive-