Professional Documents
Culture Documents
www.elsevier.com/locate/ress
PII: S0951-8320(15)00272-0
DOI: http://dx.doi.org/10.1016/j.ress.2015.09.008
Reference: RESS5408
To appear in: Reliability Engineering and System Safety
Received date: 26 March 2015
Revised date: 1 September 2015
Accepted date: 10 September 2015
Cite this article as: Enze Zhang and Qingwei Chen, Multi-objective reliability
redundancy allocation in an interval environment using particle swarm
o p t i m i z a t i o n , Reliability Engineering and System Safety,
http://dx.doi.org/10.1016/j.ress.2015.09.008
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
1
School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China
*Corresponding author. Tel.: +8615850577123.
E-mail addresses: yzzez8986@gmail.com (E. Zhang).
Abstract
Most of the existing works addressing reliability redundancy allocation problems are based on the assumption of
fixed reliabilities of components. In real-life situations, however, the reliabilities of individual components may be
imprecise, most often given as intervals, under different operating or environmental conditions. This paper deals
with reliability redundancy allocation problems modeled in an interval environment. An interval multi-objective
optimization problem is formulated from the original crisp one, where system reliability and cost are
simultaneously considered. To render the multi-objective particle swarm optimization (MOPSO) algorithm capable
of dealing with interval multi-objective optimization problems, a dominance relation for interval-valued functions
is defined with the help of our newly proposed order relations of interval-valued numbers. Then, the crowding
distance is extended to the multi-objective interval-valued case. Finally, the effectiveness of the proposed approach
has been demonstrated through two numerical examples and a case study of supervisory control and data
Keywords: Reliability optimization; Multi-objective optimization; Particle swarm optimization; Interval number
1. Introduction
The utilization of redundancy plays an important role in enhancing the reliability of a system.
The redundancy allocation problem (RAP) involves the selection of components and a
system reliability, cost and weight, given certain design constraints. The incorporation of
redundant components improves system reliability, but can also increase system cost, weight, etc.
Thus, a RAP frequently encounters a trade-off between maximization of system reliability and
Traditionally, the RAP has been solved as a single objective optimization problem with the
goal to maximize the system reliability subject to several constraints such as cost, weight, etc.
Various methodologies have been proposed to handle it, e.g., dynamic programming [1], integer
2
programming [2], column generation [3], branch and bound [4], and heuristic/meta-heuristic
approaches such as genetic algorithm [5], variable neighborhood search [6], bacterial-inspired
evolutionary algorithm [7], bee colony algorithm [8], Tabu search [9], swarm optimization [10-13],
with system design, which makes the multi-objective formulations of the reliability optimization
problem quite natural. Multi-objective approaches for the RAP can be found in [15-22].
Khalili-Damghani et al. [15] recently proposed a decision support system for solving
variant of the non-dominated sorting genetic algorithm (NSGA-II) to solve a novel mathematical
model for multi-objective RAPs. Zhang et al. [22] proposed a practical approach combining
In spite of the existence of diverse studies that address reliability optimization problems,
most of the research studies are based on the assumption of fixed reliabilities of components. In
real-life situations, however, the reliabilities of individual components are often imprecise under
different operating and environmental conditions. The causes may be improper storage facilities,
the vagueness of human judgment and other factors relating to the environment. Moreover, the
components with exactly identical reliabilities, and thus the issue is subject to uncertainty. To deal
programming [25-27], interval programming [31-33] and robust optimization [28,29] are
frequently employed. Actually, reliability optimization incorporating uncertainty has become the
subject of many research efforts in the area of reliability engineering over the last decade. Soltani
[34] presented a comprehensive review on reliability optimization with both deterministic and
non-deterministic parameters.
of the parameters is required to be known in priori. Unfortunately, it is not an easy task to grasp
this information in real engineering applications. An alternative solution to this problem is to use
3
the extent of tolerance or a region the parameter can possibly take [35,36]. Some researchers have
(see, e.g., [31-33,37-45]). In the early years, the problem was formulated by Yokota et al. [37,38]
and Taguchi et al. [39,40] as a nonlinear integer programming problem and solved by using a GA.
Gupta et al. [33] formulated the RAP with interval reliabilities as an unconstrained integer
programming problem with interval coefficients and developed an advanced GA for solving it.
Bhunia et al. [32] solved chance constrained reliability stochastic optimization problem with
resource constraints considering the reliability of each component as an interval number. Sahoo et
al. [31] formulated the problem as constrained multi-objective optimization problems with the
help of their proposed order relations of interval-valued numbers, and then converted the same
into unconstrained single objective ones. Feizollahi et al. [41] proposed a robust deviation
framework to deal with the RAP with interval component reliabilities. More recently, Soltani et al.
[42] presented a redundancy allocation model with choices of a redundancy strategy and
component type and solved it using an interval programming approach. Sadjadi et al. [43]
considered a RAP with choices of active and cold strategy where the components time to failure
follows an Erlang distribution and formulated it through a Min-Max regret method. Roy et al. [44]
considered a multi-objective RAP with both fixed and interval-valued system parameters using
entropy as a constraint for the system stability. Despite their efforts, all these studies fail to find
the trade-off front of the problem since the problem is either formulated as or converted into a
single objective optimization problem, and a multi-objective optimization approach for obtaining
allocation problems, modeled in an interval environment. Our motivation is to handle the interval
multi-objective case as-is and apply the particle swarm optimization to obtain the tradeoff curve.
formulated and then particle swarm optimization has been applied to solve the problem under a
number of constraints. To deal with imprecision, the imprecise Pareto dominance is defined,
which will be used to compare imprecise individuals. Then the crowding distance measure is
extended to handle imprecise objective functions. Finally, three numerical examples are solved
4
and the trade-off relationship between reliability and cost performance is analyzed.
The rest of the paper is organized as follows. A general interval multi-objective optimization
problem is introduced in Section 2. Section 3 describes the formulation of the interval RAP while
the proposed algorithm for solving the interval optimization problem is discussed in detail in
Section 4. Two numerical examples and a case study are presented in Section 5 to demonstrate the
effectiveness of the proposed algorithm. Finally, the conclusions and suggestions for future
Very often real-world problems have several conflicting objectives. Even though many of
them can be reduced to a matter of a single objective, it may not adequately represent the problem
being faced. Considering multiple objectives often gives better ideas of the task. If so, there is a
vector of objectives involving M ( 2) conflicting objective functions that must be traded off in
some way. Multi-objective optimization is concerned with the minimization of y that can be the
where the decision vectors x = (x1 , x2 ,, xn )T belong to the feasible region S R which is
n
Note that because y (x) is a vector, if any of the components of y (x) are competing, there
determine what is considered as an optimal solution, and this criterion is Pareto dominance.
said to dominate x 2 (denoted by x1 x 2 ), if and only if both of the following conditions are
true:
yi (x1 ) yi (x2 )
The above relationship is clearly suitable for precise objective values, but when it comes to
objective functions that are imprecise, represented as multidimensional intervals [50] (see Fig. 1),
the normal Pareto dominance relation cease to work and need to be extended, as will be discussed
later. In an interval environment, the original problem (1) is converted into an interval
where c = (c1 ,c2 ,,cK )T are the interval-valued parameter; ckL and ckR are the lower and
upper bounds of the kth component; yi (x) = [ yiL (x), yiR (x)] (i = 1,2,, M) are the ith
y2 y2
y
y" max(y ) 1R
y2 R
y y" y
y
min(y ) 1L
y' y2 L
y'
y1 y1
3. Problem formulation
The RAP pertains to a system of n subsystems arranged in series. The ith subsystem
cost, weight and other characteristics. The use of redundant components improves system
reliability, but also increases system cost and weight. The problem arises then regarding how to
illustrated in Fig. 2.
6
1 1 1
2 2 2
x1 x2 xn
3.1. Assumptions
The basic assumptions for the RAP in an interval environment are as follows:
All components are assumed to have binary states, i.e. operating and failure.
optimization problem, in which the reliabilities and costs of components are interval-valued.
Assuming that the component type of the ith subsystem is identical, the system reliability RS
equals to the product of the interval-valued reliability of all subsystems (see [31-33,35,36] for the
where RSL (x) [1 (1 riL ) xi ] and RSR (x) [1 (1 riR ) xi ] . The problem is to determine the
number of redundant components xi (i = 1,2,,n) that will maximize the system reliability and
minimize the system cost under given constraints. The mathematical formulation of the problem is
given below:
n
max RSL (x), RSR (x)
i=1
where [ RSL , RSR ] and [CSL , CSR ] are the interval-valued system reliability and system cost
7
respectively.
4. Proposed approach
To develop an interval MOPSO algorithm for Problem (4), modifications are necessary in
those steps where imprecise solutions are compared. In the following subsections a dominance
relation for interval-valued objective functions is defined. Then, the crowding distance metric is
As previously mentioned, the Pareto dominance relation is not suitable for interval-valued
objective functions. Thus, a dominance relation therefore has to be defined for comparison of
individuals. Extending Pareto dominance to the interval-valued case, we face the difficulty of
comparison between two interval numbers. Regarding this, a new partial order relation that
are the left and right limits on the real line R , respectively. If aL aR , then A [a, a] is a real
number. Interval A can also be denoted by A = ac ,aw , where ac and aw are the mid-point
1 1
and half-width of interval A , i.e., ac (aL aR ) and aw (aR aL ) .
2 2
An extensive research on comparing and ranking two interval numbers can be found in [36]
which gives two approaches of comparing two interval numbers. The first one describes an
optimistic decision makers preference index. The second one defines strict and fuzzy preference
ordering from a pessimistic decision makers point of view. This approach is somewhat subjective
since the degree of pessimism of decision makers should be specified beforehand. To overcome
this issue, we have proposed a general definition of order relations without extra knowledge about
Clearly the order relation IM is reflexive and transitive but not symmetric.
Applying IM , a relation IP extending the standard Pareto dominance is proposed for the
(5)
To produce well-distributed Pareto fronts, the global best position of each particle is selected
from the external repository with respect to the diversity of nondominated solutions. IP-MOPSO
employs the crowding distance to estimate the diversity of nondominated solutions. However, the
crowding distance first introduced in [46] is not suitable for interval-valued objective functions,
and thus, a novel crowding distance extending the previous one need to be developed.
Considering two optimal solutions xi and x j , the distance between them can be calculated
as follows:
M
d( f m (x i ), f m (x j ))
D(x ,x ) = m=1
, (6)
i j
V (x i ) +V (x j ) +1
where M indicates the dimension of objective functions; f m (xi ) [ fmL (xi ), f mR (xi )] and
mth objective function; d ( f m (xi ), f m (x j )) denotes the distance between f m (xi ) and f m (x j )
Assuming the two nearest points x j and x k , the crowding distance of xi , denoted by
It is worth noting that the solutions that lie in the boundary of each dimension of the objective
It can be observed from Eq. (8) that if the interval objectives represented as intervals regress
to points, then the distance between xi and x j will take the following form:
M
D(x i ,x j ) = | f m (x i ) - f m (x j ) | (9)
m=1
1 M
CD(xi ) | f m (x j ) f m (x k ) |
2 m 1
(10)
which is consistent with the normal one for the non-interval case. From this point of view, the
crowding distance introduced in [46] is a special case of the one defined here.
Particle swarm optimization (PSO), first introduced by Kenney and Eberhart in 1995 [47], is a
stochastic global optimization technique inspired by the paradigm of birds flocking. In the PSO, a
swarm consists of a set of particles that fly around in a multi-dimensional search space. Each
particle represents a potential solution of an optimization problem. During the flight, each particle
adjusts its position according to the experience of itself and other neighboring particles, making
use of the best position they encounter. Considering the i th particle in the swarm at iteration t ,
the personal best position, known as pbesti , is the best position found so far by itself, while the
global best position, known as gbest , is the global best position found so far by neighbors of this
10
particle, and its position and velocity are denoted by xit and vit , respectively. Then, the position
and velocity of this particle at iteration t 1 can be expressed using the following equations:
where c1 and c2 are the acceleration coefficients, which control the influence of gbest and
pbesti on the search process, r1 and r2 are random numbers between 0 and 1, and w is the
inertia weight used to control the impact of the previous velocities on the current one, influencing
It is found that the performance of a PSO algorithm could be improved by linearly varying
the inertia weight w [48]. We adopt in IP-MOPSO the time-varying inertia weight proposed by
where w1 and w2 are the initial and final values of the inertia weight respectively, t is the
Besides using this inertia weight scheme, time-varying acceleration coefficients are applied
with the idea of balancing the ability of exploration and exploitation. This can be carried out by
linearly changing the coefficients c1 and c2 through the time as suggested in [50]:
t
c1 (c1 f c1i ) c1i (13)
Tmax
t
c2 (c2 f c2i ) c2i (14)
Tmax
where c1i and c1 f are the initial and final values of c1 , c2i and c2 f are the initial and final
values of c2 . Here, t and Tmax have the same definition as in Eq. (12).
constraint-handling scheme needs to be incorporated to deal with the constraints. The Big-M
11
penalty method introduced in [33] is considered in the IP-MOPSO algorithm, which converts the
one by penalizing the unfeasible solutions a large positive number, say M , which can be written
in an interval form [M , M ] .
Maximize [f L , f R ]
(15)
s.t. g j (x) 0, j = 1,2,,m.
Let S = {x : g j (x) 0, j = 1,2,,m} be the feasible space, then the transformed optimization
problem is as follows:
[0,0], if x S ;
where ( x)
f (x) [ M , M ], otherwise.
The proposed algorithm, IP-MOPSO, which adopts the imprecise Pareto relation and
Step 1: Initialize the population of particles. The personal best position is set as the particle
itself. Evaluate each of the particles using Eq. (16) and store the nondominated ones in the
Step 2: Randomly select the global best position from the top 10% of solutions in Rep
Step 3: Update the velocity and position of the particles according to Eq. (11). A mutation
operator with action range covered narrowed over time is adopted to prevent a premature
Step 4: Evaluate each of the particles and store the nondominated ones in the external
Step 5: Update the personal best position of the particles. For every particle, between its
12
current position and the previous personal best position, the dominating one, if any, will be
the new personal best position, and any position randomly otherwise.
Step 6: Update the external repository based on the crowding distance suitable for
interval-valued objectives. This update consists of inserting all the currently nondominated
solutions into the repository and eliminating the dominated ones. Since the size of the
repository is limited, whenever it gets full, those particles located in less populated regions
of objective space, i.e., N r solutions with greater crowding distances, are kept in the
repository.
Step 7: Increase the iteration counter. If the maximum number of cycles is reached, the
Algorithm IP-MOPSO
1: t 0
2: FOR i 1 to nPop
3: ( xi , gbesti , pbesti ) Initialize( ); //*initialize the population Pop0 *//
4: ENDFOR
5: F ( Pop0 ) Evaluate( Pop0 ) //*Evaluate each of the particles in Pop0 *//
6: Rep GetNDParticles( Pop0 ) //*Store the nondominated particles in the repository using IP
*//
7: WHILE t Tmax DO
8: FOR i 1 to nPop
9: gbesti GetGbest( )
10: pbesti GetPbest( )
11: xi UpParticle( xi , gbesti , pbesti )
12: xi Mutate( xi )
13: END FOR
14: F ( Popt 1 ) Evaluate( Popt 1 )
15: NDPaticles GetNDParticles( Popt 1 )
16: Rep GetNDParticles(Rep NDPaticles) //*Update the external repository*//
17: IF | Rep | nRep THEN Rep Prune( Rep) //* | Rep | : number of particles in Rep *//
18: t t 1
19: END WHILE
13
5. Numerical examples
Three examples are considered to illustrate and validate the performance of the proposed
techniques for solving multi-objective RAPs with interval-valued reliabilities of components. The
values of the parameters considered in the third example are selected from a real case study.
Example 1: An example problem taken from [33] is solved in the first example, which
consists of five subsystems connected in series. For each subsystem, the component type available
i 1
5
Minimize CSL , CSR CiL , CiR xi exp( xi / 4)
i 1
s.t. (17)
5
g1 ( x) Pi xi2 b1 0
i 1
5
g 2 ( x) Wi xi exp( xi / 4) b2 0
i 1
and xi (i 1, 2,3, 4,5) being a nonnegative integer; where the values of Ci , Pi , Wi , b1 and b2
Table 1
Data for example 1
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]
Ci [6,8] [5,8] [3,6] [6,9] [3,6]
Pi 1 2 3 4 2
Wi 7 8 8 6 9
b1 110, b2 200.
The problem was solved using the developed algorithm, IP-MOPSO, with a population size
of 30, an archive size of 15 and 50 iterations. The values of c1 , c2 , w1 and w2 have been taken
as 2.5, 0.5, 0.9 and 0.4 respectively. Table 2 presents 15 of the solutions found with their
respective system reliability and cost. The range of system reliability is from [0.2657, 0.3181] to
[0.8839, 0.9132]. Correspondingly, system cost ranges from [52.5326, 84.5089] to [105.9448,
14
168.7731]. These ranges assure the extent of the generated Pareto front. This Pareto front provides
decision-makers different options for system implementation. The results obtained are compared
From Table 2 and 3, we observe that solution no. 3# is the same as solution no. 2 obtained by
IP-MOPSO, which is indicated in italic type. Solution no. 1# is obviously dominated by solution
no. 4 and 13 shown in boldface. Solution no. 2# is dominated by solution no. 13. Fig. 3 further
shows solutions obtained by IP-MOPSO and from [33], where it can be seen that IP-MOPSO
performs better.
Table 2
Results obtained by IP-MOPSO of Example 1
Solution no. x's Reliability Cost
1 [1,1,1,1,1] [0.2657,0.3181] [52.5326,84.5089]
2 [3,2,2,3,3] [0.8839,0.9132] [105.9448,168.7731]
3 [1,1,1,1,2] [0.3348,0.3945] [56.6267,92.6971]
4 [2,1,1,2,2] [0.5596,0.6238] [73.0030,115.8969]
5 [2,2,2,3,3] [0.8502,0.8888] [97.1351,157.0269]
6 [2,2,2,3,2] [0.8069,0.8494] [92.7303,148.2172]
7 [1,1,1,2,2] [0.4587,0.5286] [64.8148,104.9794]
8 [2,1,2,2,2] [0.6324,0.6799] [77.0971,124.0851]
9 [2,2,2,2,2] [0.7336,0.7819] [83.9206,135.0027]
10 [2,2,3,3,3] [0.8629,0.8954] [101.5399,165.8365]
11 [1,1,2,1,2] [0.3784,0.4300] [60.7208,100.8853]
12 [2,2,2,2,3] [0.7729,0.8182] [88.3254,143.8124]
13 [1,1,2,2,2] [0.5184,0.5762] [68.9089,113.1676]
14 [2,1,1,1,2] [0.4085,0.4655] [64.8148,103.6147]
15 [2,1,2,2,3] [0.6663,0.7115] [81.5019,132.8948]
Table 3
Results obtained in [33] of Example 1
Solution no. x's Reliability Cost
1# [2,1,2,1,3] [0.4864,0.5310] [73.3183,120.6125]
2# [1,2,2,1,3] [0.4625,0.5175] [71.9491,120.6125]
3# [3,2,2,3,3] [0.8839,0.9132] [105.9448,168.7731]
15
180
160
140
120
Cost
100
80
60
40
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Reliability
Example 2: The second example considers a system consisting of three subsystems, with an
option of five, four and five types of components in each subsystem, respectively. The maximum
s.t. (18)
mi
xij {0,1,2,,nmax,i }
where xij is the decision variable denoting the number of the jth components used in
subsystem i ; mi is the number of available component types for each subsystem i ; nmax,i is
the user-defined maximum number of components used in subsystem i which is eight in this
Table 4
Component choices of example 2
Component type j Subsystem i
1 2 3
rij cij rij cij rij cij
1 [0.93,0.95] [8,10] [0.96,0.98] [11,13] [0.95,0.97] [9,11]
2 [0.90,0.92] [5,7] [0.85,0.87] [2,4] [0.88,0.90] [5,7]
3 [0.88,0.90] [5,7] [0.69,0.71] [1,3] [0.71,0.73] [3,5]
4 [0.74,0.76] [2,4] [0.65,0.67] [1,3] [0.70,0.72] [2,4]
5 [0.71,0.73] [1,3] [0.66,0.68] [1,3]
16
200
180
160
140
120
Cost
100
80
60
40
20
0
0.4 0.5 0.6 0.7 0.8 0.9 1
Reliability
180
C
160
140 System reliability
maximization
Average cost
120
100
80
60
System cost
40 minimization
B
20 A
0
0.4 0.5 0.6 0.7 0.8 0.9 1
Average reliability
For this case, IP-MOPSO was run considering a population size of 50, and 100 generations.
Fig. 4 shows a sample population plot of IP-MOPSO, where it can be observed that the solutions
are quite diverse and form a broad Pareto optimal front. To better illustrate the trade-off
relationship between reliabilities and costs, Fig. 5 presents the Pareto optimal front generated by
trade-off points with the mid-point of intervals representing objective functions. Each point on this
frontier represents an optimal design configuration with different system reliabilities and costs.
represents the design configuration of reliability maximization, which leads to the highest cost.
Moreover, there is one obvious turning point on the Pareto frontier, point B as labeled. As
observed, a high increment of reliability is gained if the selection of solutions passes from point
A to point B than when it changes from point B to C . This means that passing from
17
solution A to B brings a small increment in the system cost but a high improvement in the
system reliability.
Table 5
Solutions corresponding to three trade-off points of Example 2
Sol. Points x's Reliability Cost
A [0,0,0,0,1,0,0,1,0,0,0,0,0,1] [0.323334000, 0.352444000] [3,9]
B [0,0,0,2,2,0,2,0,0,1,0,0,0,3] [0.960482384, 0.969989817] [15,35]
C [0,6,1,1,0,5,1,1,0,3,5,0,0,0] [0.999999961, 0.999999993] [147,193]
A decision-maker can pick up any point from the Pareto front according to their specific
design criteria or interest. Once a design point is selected, the system configuration behind it can
be obtained. Table 5 illustrates the solutions representing the design configurations corresponding
Now a natural question comes: Does the imprecision Pareto dominance based MOPSO
perform better than the standard distribution assuming one? To investigate this point, we choose
imprecision by assuming a uniform distribution inside the objective vector. The expectation value
of the objective vector is used for optimization and thus the algorithm becomes a normal one.
Both algorithms were run for 100 generations with a population size of 50. Results were
approximation performance of a nondominated set X , which has also interval form [50]. It
Spacing metric ( SP ) : To evaluate the distribution of solutions throughout the found Pareto
front, we introduce a metric that is an extension of the spacing metric originally proposed in [51].
1 n
D Di
2
SP (19)
n 1 i 1
solutions; D is the mean value of all Di . A value of zero for this measure indicates that all the
18
Two-set coverage metric (C ) : In order to evaluate the closeness of the Pareto front found to
the true one that is unknown in advance, we extend the two-set coverage (C ) metric originally
proposed in [52] to the imprecise case. This metric takes a pair of nondominated solution sets X
|{y Y | $x X , x IP y x = y}|
C( X ,Y ) = (20)
|Y |
where | | denotes the number of members of the solution set. C ( X , Y ) 1 indicates that all
Imprecision metric ( I ) : Another performance considered for the imprecise case is the
amount of uncertainty in a population. This might be measured using the I metric [50] as the
Table 6
Comparison results of the metric C, SP and I.
C SP I
C (IP-*, DA-*) C(DA-*, IP-*) Mean Std. Mean Std.
IP-MOPSO 0.2613 0.0157 0.0036 0.1812 0.0038
DA-MOPSO 0.4706 0.0243 0.0025 0.3709 0.0072
represents MOPSO
0.8
0.7
0.6
S
0.5
0.4
Best case: IP-MOPSO
Best case: DA-MOPSO
0.3 Worsst case: IP-MOPSO
Worst case: DA-MOPSO
0.2
0 20 40 60 80 100
Iteration
Fig. 6. Comparison results of S metric
Detailed plots of S metric over the optimization run as shown in Fig. 6 indicate that
19
IP-MOPSO seems to have a better convergence performance than DA-MOPSO. The C metric
values in Table 6 emphasize this observation since more than 26.1% solutions generated by
IP-MOPSO are dominated by those of DA-MOPSO. However, nearly 47.1% of the solutions
obtained by DA-MOPSO are dominated by those of IP-MOPSO. Besides, the mean and standard
deviation of the SP and I results over 30 independent simulation runs are given in Table 6.
These results indicate that the IP-MOPSO algorithm performs better than DA-MOPSO in terms of
The conclusion to draw from all these observations is that the imprecise Pareto dominance
(DA-MOPSO).
Next, the proposed approach for solving interval-valued multi-objective RAP is illustrated
The SCADA system considered has three main sub-systems, i.e., modems, FEPs, and servers,
which work serially. Modems are responsible for communication between stations and the main
control center. FEPs collect data from stations via a proper protocol. Servers are used for acquiring
In each subsystem, five types of redundant components can be used. The maximum number
s.t. (18)
mi
xij {0,1,2,,nmax,i }
where xij is the decision variable denoting the number of the jth components used in
subsystem i ; mi is the number of available component types for each subsystem i ; nmax,i is
the user-defined maximum number of components used in subsystem i which is five in this case.
20
The reliabilities and costs of all the components are selected from the case study in [19], and
Table 7
Component choices of example 3
Component Subsystem i
type j 1 (Server) 2 (FEP) 3 (Modem)
rij cij rij cij rij cij
1 [0.784,0.816] [98,102] [0.931,0.969] [490,510] [0.931,0.969] [1960,2040]
2 [0.882,0.918] [196,204] [0.882,0.918] [441,459] [0.588,0.612] [1764,1836]
3 [0.686,0.714] [78.4,81.6] [0.784,0.816] [392,408] [0.882,0.918] [2050,2150]
4 [0.735,0.765] [83.3,86.7] [0.735,0.765] [343,357] [0.931,0.969] [2131.5,2218.5]
5 [0.833,0.867] [147,153] [0.9506,0.9894] [514.5,535.5] [0.833,0.867] [2009,2091]
corresponding type is selected as the encoded elements. Several elements compose a particle
representing a candidate solution of the RAP problem. The structure of a particle is illustrated in
Fig. 7, where xij represents the number of redundant component of type j in subsystem i.
Each particle contains a 15-bit integer coded string, which consists of 3 subsystems, with an
option of 5 types of components in each subsystem. This integer-encoding scheme can achieve the
mapping from the particle representation to the RAP to be solved in a convenient way.
0.9
0.8
Reliability
0.7
0.6
0.5
0.4
0.3
2000 4000 6000 8000 10000 12000 14000 16000
Cost
0.9
0.8
Reliability
0.7
0.6
0.5
0.4
0.3
2000 4000 6000 8000 10000 12000 14000 16000
Cost
Fig. 9. Pareto optimal front obtained by IP*-MOPSO
The Pareto optimal front generated by IP-MOPSO is shown in Fig. 8, where it can be seen
that the solutions obtained are quite diverse and form a broad optimal set. The system reliability
value varies from [0.3162, 0.3566] to [0.99994403, 0.99998972], with the corresponding total cost
varying from [2325.6, 2234.4] to [13254, 13796]. To further show the superiority of the proposed
order relation for interval-valued numbers and the Pareto dominance relation, a well-known Pareto
relation > IP proposed in [50], adopting the order relation > IN , is used for comparison purpose.
To tell them apart, the MOPSO algorithm adopting the Pareto relation > IP is named IP*-MOPSO,
with all the other parts of the algorithm exactly the same as in IP-MOPSO. Fig. 8 shows the Pareto
optimal front obtained by IP*-MOPSO. We observe from Fig. 8 and 9 that a wide range of values
is covered by solutions of both algorithms. However, it is evident that the solutions obtained using
IP-MOPSO are better distributed through the tradeoff curve, which means that IP-MOPSO
0.95
Reliability
0.9
0.85
0.8
2000 4000 6000 8000 10000 12000 14000
Cost
Fig. 10. Pareto front with a lower bound of 0.8 for reliability
22
Table 8
Example system design configurations from the knee region
Sol. No. System configuration diagram Reliability Cost
2 1
5
19 2 1 [0.9956, 0.9993] [7497,7803]
5
2 1
1
2 2
5
13 2 2 [0.9971,0.9998] [7673.4,7986.6]
5
2 2
3
2
4
2
5 1
43 5 [0.9951, 0.9990] [6639.5,6910.5]
5 1
5
5
5
A lower bound of 0.8 for reliability is set to identify the solutions more likely to be chosen by
decision-makers, which are plotted in Fig. 10. To further find the good compromises, solutions from
the knee region are considered. The knee region is formed by the most interesting solutions of the
Pareto front, i.e., those where a small improvement of one objective would lead to a large deterioration
in at least one other objective [18,25]. In this case, solutions from the knee region are those where a
small improvement of reliability would lead to a huge increase of cost. Table 8 illustrates three
solutions from this region representing system design configurations, with their respective
6. Conclusion
component has been discussed here, where the reliability and cost of a system are simultaneously
considered and properly traded off. A multi-objective particle swarm optimization algorithm,
which adopts the imprecise Pareto relation and extended crowding distance, is proposed for
solving it. A Pareto frontier can be obtained, which captures all possible types of system design
under certain design criteria and conditions and thus provides a decision maker a full set of design
configurations for implementation. For future research, this paper could be extended in four
directions:
23
1. When the number of objective functions is more than two, the proposed crowding distance
may result in losing spatial information. Thus, a first future research direction is to propose a
crowding distance which can reflect the spatial information well and works better when the
2. It should be noted that the general idea presented here could also be applied to other
system structures such as k-out-of-n, circular, complex, and bridge. Hence, another future research
However, decision makers, generally risk-averse, prefer designs with a high reliability assured by
strategy can be considered. Also, the model can be extended to allow component mixing that
Acknowledgement
The authors would like to thank the editor, the associate editor and the anonymous reviewers
for their insightful comments and suggestions that help us to improve the quality of the paper. This
work is supported by the National Natural Science Foundation of China under Grant No.
61333008.
References
[1] Yalaoui A, Chtelet E, Chu C. A new dynamic programming method for reliability & redundancy allocation
in a parallel-series system. IEEE Trans Reliab 2005;54(2):254-61.
[2] Billionnet A. Redundancy allocation for series-parallel systems using integer linear programming. IEEE
Trans Reliab 2008;57:507-16.
[3] Zia L, Coit DW. Redundancy allocation for series-parallel systems using a column generation approach.
IEEE Trans Reliab 2010;59:706-17.
[4] Ha C, Kuo W. Reliability redundancy allocation: An improved realization for nonconvex nonlinear
programming problems. Eur J Oper Res 2006;171(1):24-38.
[5] Ardakan MA, Hamadani AZ. Reliability optimization of series-parallel systems with mixed redundancy
strategy in subsystems. Reliab Eng Syst Saf 2014;130:132-9.
[6] Liang YC, Chen YC. Redundancy allocation of series-parallel systems using a variable neighborhood search
algorithm. Reliab Eng Syst Saf 2007;92(3):323-31.
[7] Hsieh TJ. Hierarchical redundancy allocation for multi-level reliability systems employing a
bacterial-inspired evolutionary algorithm. Inf Sci 2014;288:174-93.
[8] Yeh W C, Hsieh TJ. Solving reliability redundancy allocation problems using an artificial bee colony
algorithm. Comput Oper Res 2011;38(11):1465-73.
24
[9] Ouzineb M, Nourelfath M, Gendreau M. Tabu search for the redundancy allocation problem of
homogenous series-parallel multi-state systems. Reliab Eng Syst Saf 2008;93(8):1257-72.
[10] Yeh WC. Orthogonal simplified swarm optimization for the series-parallel redundancy allocation
problem with a mix of components. Knowledge-Based Systems 2014;64:1-12.
[11] Huang CL. A particle-based simplified swarm optimization algorithm for reliability redundancy allocation
problems. Reliab Eng Syst Saf 2015;142:221-30.
[12] Wang Y, Li L. A PSO algorithm for constrained redundancy allocation in multi-state systems with bridge
topology. Comput Ind Eng 2014;68:13-22.
[13] Kong X, Gao L, Ouyang H, Li S. Solving the redundancy allocation problem with multiple strategy choices
using a new simplified particle swarm optimization. Reliab Eng Syst Saf 2015;144:147-58.
[14] Kanagaraj G, Ponnambalam SG, Jawahar N. A hybrid cuckoo search and genetic algorithm for
reliability-redundancy allocation problems. Comput Ind Eng 2013;66(4):1115-24.
[15] Khalili-Damghani K, Abtahi AR, Tavana M. A Decision Support System for Solving MultiObjective
Redundancy Allocation Problems. Qual Reliab Eng Int 2014;30(8):1249-62.
[16] Cao D, Murat A, Chinnam RB. Efficient exact optimization of multi-objective redundancy allocation
problems in series-parallel systems. Reliab Eng Syst Saf 2013;111:154-63.
[17] Zaretalab A, Hajipour V, Sharifi M, Shahriari MR. A knowledge-based archive multi-objective simulated
annealing algorithm to optimize series-parallel system with choice of redundancy strategies. Comput Ind Eng
2015;80:33-44.
[18] Dolatshahi-Zand A, Khalili-Damghani K. Design of SCADA water resource management control center
by a bi-objective redundancy allocation problem and particle swarm optimization. Reliab Eng Syst Saf
2015;133:11-21.
[19] Li ZJ, Liao HT, Coit DW. A two-stage approach for multi-objective decision making with applications to
system reliability optimization. Reliab Eng Syst Saf 2009;94:1585-92.
[20] Safari J. Multi-objective reliability optimization of series-parallel systems with a choice of redundancy
strategies. Reliab Eng Syst Saf 2012;108:10-20.
[21] Khalili-Damghani K, Abtahi AR, Tavana M. A new multi-objective particle swarm optimization method
for solving reliability redundancy allocation problems. Reliab Eng Syst Saf 2013;111(2):58-75.
[22] Zhang EZ, Wu YF, Chen QW. A practical approach for solving multi-objective reliability redundancy
allocation problems using extended bare-bones particle swarm optimization. Reliab Eng Syst Saf
2014;127:65-76.
[23] Zhao R, Liu B. Stochastic programming models for general redundancy-optimization problems. IEEE
Trans Reliab 2003;52(2):181-91.
[24] Tekiner-Mogulkoc H, Coit D W. System reliability optimization considering uncertainty: Minimization
of the coefficient of variation for series-parallel systems. IEEE Trans Reliab 2011;60(3):667-74.
[25] Wang S, Watada J. Modelling redundancy allocation for a fuzzy random parallel-series system. J
Comput Appl Math 2009;232(2):539-57.
[26] Garg H, Rani M, Sharma SP, Vishwakarma Y. Bi-objective optimization of the reliability-redundancy
allocation problem for series-parallel system. J Manuf Syst 2014;33(3):335-47.
[27] Soltani R, Sadjadi SJ. Reliability optimization through robust redundancy allocation models with
choice of component type under fuzziness. Proceedings of the Institution of Mechanical Engineers, Part
O: Journal of Risk and Reliability 2014;228(5):449-59.
[28] Feizollahi MJ, Ahmed S, Modarres M. The robust redundancy allocation problem in series-parallel
systems with budgeted uncertainty. IEEE Trans Reliab 2014;63(1):239-50.
[29] Feizollahi MJ, Soltani R, Feyzollahi H. The robust cold standby redundancy allocation in series-parallel
systems with budgeted uncertainty. IEEE Trans Reliab 2015;64(2):799-806.
[30] Do DM, Gao W, Song C, Tangramvong S. Dynamic analysis and reliability assessment of structures
with uncertain-but-bounded parameters under stochastic process excitations. Reliab Eng Syst Saf
2014;132:46-59.
[31] Sahoo L, Bhunia AK, Kapur PK. Genetic algorithm based multi-objective reliability optimization in
interval environment. Comput Ind Eng 2012;62(1):152-60.
[32] Bhunia AK, Sahoo L, Roy D. Reliability stochastic optimization for a series system wi th interval
component reliability via genetic algorithm. Appl Math Comput 2010;216(3):929-39.
[33] Gupta RK, Bhunia AK, Roy D. A GA based penalty function technique for solving constrained
redundancy allocation problem of series system with interval valued reliability of components. J
Comput Appl Math, 2009;232(2): 275-84.
[34] Soltani R. Reliability optimization of binary state non-repairable systems: A state of the art survey. Int J
Ind Eng Comput 2014;5(3):339-64.
[35] Sengupta A, Pal TK, Chakraborty D. Interpretation of inequality constraints involving interval
coefficients and a solution to interval linear programming. Fuzzy Sets Syst 2001 ;119(1):129-38.
25
[36] Sengupta A, Pal TK. On comparing interval numbers. Eur J Oper Res 2000;127(1):28-43.
[37] Yokota T, Gen M, Taguchi T, et al. A method for interval 01 nonlinear programming problem using a
genetic algorithm. Comput Ind Eng 1995;29(1):531-5.
[38] Yokota T, Gen M, Li Y, Kim CE. A genetic algorithm for interval nonlinear integer programming
problem. Comput Ind Eng 1996;31(3):913-7.
[39] Taguchi T, Ida K, Gen M. Method for solving nonlinear goal programming with interval coeffic ients
using genetic algorithm. Comput Ind Eng 1997;33(3):597-600.
[40] Taguchi T, Yokota T. Optimal design problem of system reliability with interval coefficient using
improved genetic algorithms. Comput Ind Eng 1999;37(1):145-9.
[41] Feizollahi MJ, Modarres M. The robust deviation redundancy allocation problem with interval
component reliabilities. IEEE Trans Reliab 2012;61(4):957-65.
[42] Soltani R, Sadjadi SJ, Tavakkoli-Moghaddam R. Interval programming for the redundancy allocation
with choices of redundancy strategy and component type under uncertainty: Erlang time to failure
distribution. Appl Math Comput 2014;244:413-21.
[43] Sadjadi SJ, Soltani R. Minimum-Maximum regret redundancy allocation with the choice of redundancy
strategy and multiple choice of component type under uncertainty. Comput Ind Eng 2015 ;79:204-13.
[44] Roy P, Mahapatra BS, Mahapatra GS, Roy PK. Entropy based region reducing genetic algorithm for
reliability redundancy allocation in interval environment. Exp Syst Appl 2014;41:6147-60.
[45] Soltani R, Sadjadi SJ, Tavakkoli-Moghaddam R. Robust cold standby redundancy allocation for
nonrepairable series-parallel systems through Min-Max regret formulation and Benders decomposition
method. Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability
2014;228(3):254-64.
[46] Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective genetic algorithm: NSGA-II.
IEEE Trans Evol Comput 2002;6(2):182-97.
[47] Kennedy J, Eberhart RC. Particle swarm optimization. In: Proceedings of the 4th IEEE International
conference on neural networks. Piscataway, NJ; 1995.p. 1942-8.
[48] Shi Y, Eberhart RC. Empirical study of particle swarm optimization. In: Proceedings of the 1999
Congress on Evolutionary Computation. Washington, DC; 1999. p. 1945-50.
[49] Ratnaweera A, Halgamuge S, Watson HC. Self-organizing hierarchical particle swarm optimizer with
time-varying acceleration coefficients. IEEE Trans Evol Comput 2004;8(3):240-55.
[50] Limbourg P, Aponte DES. An optimization algorithm for imprecise multi-objective problem functions.
In: Proceedings of the 2005 IEEE Congress on Evolutionary Computation. Edinburgh; 2005. p. 459-66.
[51] Schott JR. Fault tolerant design using single and multicriteria genetic algorithm optimization.
Cambridge: Massachusetts Institute of Technology, 1995.
[52] Zitzler E, Deb K, Thiele L. Comparison of multiobjective evolutionary algorithms: Empirical results.
Evol Comput 2000;8(2):173-95.
Highlights
We model the reliability redundancy allocation problem in an interval
environment.
We apply the particle swarm optimization directly on the interval values.
A dominance relation for interval-valued multi-objective functions is defined.
The crowding distance metric is extended to handle imprecise objective
functions.