Applied Mathematics and Computation 182 (2006) 1556–1567

www.elsevier.com/locate/amc

IAs based approach for reliability

redundancy allocation problems
Ta-Cheng Chen
Department of Information Management, National Formosa University, 64 Wenhua Road, Huwei, Yulin 632, Taiwan

Abstract

Nonlinearly mixed-integer reliability design problems are investigated in this paper where both the number of redun-
dancy components and the corresponding component reliability in each subsystem are to be decided simultaneously so
as to maximize the reliability of system. The reliability design problems have been studied in the literature for decades,
usually using mathematical programming or heuristic/metaheuristic optimization approaches. The difficulties confronted
for both methodologies are to maintain feasibility with respect to three nonlinear constraints, namely, cost, weight and
volume related constraints. A penalty guided artificial immune algorithm is presented for solving such mixed-integer reli-
ability design problems. It can search over promising feasible and infeasible regions to find the feasible optimal/near opti-
mal solution effectively and efficiently. Numerical examples indicate that the proposed approach performs well for the
reliability-redundant allocation design problems considered in this paper. As reported, solutions obtained by the proposed
approach are as well as or better than the previously best-known solutions.
 2006 Elsevier Inc. All rights reserved.

Keywords: Reliability design; Immune algorithm; Mixed-integer nonlinear programming

1. Introduction

The system reliability optimization is very important in the real world applications and the various kinds of
systems have been studied in the literature for decades. Many designers are devoted to improving the reliabil-
ity of manufacturing systems or product components to be more competitive in the market. Generally two
major ways have been used to achieve higher system reliability. The first way is by increasing the reliability
of system components, and the second way is by using redundant components in various subsystems. In
the first way, the system reliability can be improved to some degree, but the required reliability enhancement
may be never attainable even though the most currently reliable elements are used. Using the second way is to
choose the optimal element combination and redundancy-levels; the system reliability can be also enhanced,
but the cost, weight, volume, etc. will be increased as well. It perhaps the most common problem in design-for-
reliability and has been called as a ‘‘redundancy allocation problem’’ [16]. Besides the above two approaches,

E-mail address: [email protected]

0096-3003/$ - see front matter  2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2006.05.044
 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567 1557

Nomenclature

m the number of subsystems in the system

ni the number of components in subsystem i, 1 6 i 6 m
n  (n1, n2, . . . , nm), the vector of the redundancy allocation for the system
ri the reliability of each component in subsystem i, 1 6 i 6 m
r  (r1, r2, . . . , rm), the vector of the component reliabilities for the system
qi = 1  ri, the failure probability of each component in subsystem, 1 6 i 6 m
Ri(ni) ¼ 1  qni i , the reliability of subsystem i, 1 6 i 6 m
Rs the system reliability
gi the ith constraint function
wi the weight of each component in subsystem i, 1 6 i 6 m
vi the volume of each component in subsystem i, 1 6 i 6 m
ci the cost of each component in subsystem i, 1 6 i 6 m
V the upper limit on the sum of the subsystems’ products of volume and weight
C the upper limit on the cost of the system
W the upper limit on the weight of the system

the conjunction of the two approaches and reassignment of interchangeable elements are another feasible ways
for enhancing the system reliability [9,16]. Such problems of maximizing system reliability through redun-
dancy and component reliability choices are called ‘‘reliability-redundancy allocation problem’’ [16]. These
optimization problems are very difficult but realistic situation in reliability optimization. Therefore, how to
trade of between reliability and the resources becomes a significant issue in designing a high reliability system.
Based on the ways for enhancing system reliability, the reliability design problems include the integer and
mixed-integer reliability problems. For the integer reliability problems (redundancy allocation problems), the
component redundancy allocation is to be decided while the component reliabilities are given. The major focus
of recent work in the redundancy allocation problems is on the development of heristic/metaheuristic algo-
rithms for redundancy allocation [1,2,12,22]. Only little works is directed toward exact solutions for such
problems [16]. For optimizing the system reliability of mixed-integer reliability problems (reliability-redun-
dancy allocation problems), the number of redundant components and the corresponding component reliabil-
ities are to be decided simultaneously. Few approaches were proposed for the mixed-integer reliability
problems of optimizing both the redundancy and component reliability [9,12,17,26] in literature. These system
reliability problems are either subjected to the linear constraints [8,20,21] or to the nonlinear constraints
[9,13,23,26]. Kuo and Prasad [16], Tillman et al. [24] and Yokota et al. [27] presented a good comprehensive
survey of previous works. In this paper, four mixed-integer reliability problems with multiple nonlinear con-
straints are considered and solved by using a novel metaheuristic approach. The first three example problems
include series system, series-parallel system and complex (bridge) system [9,13,17,26]. The last example prob-
lem is a gas turbine overspeed protection system [5,27].
Four reliability-redundancy allocation problems of maximizing the system reliability subject to multiple
nonlinear constraints can be stated as following nonlinearly mixed-integer programming model.
Max Rs ¼ f ðr; nÞ; ð1Þ
Subject to gðr; nÞ 6 l ð2Þ
0 6 ri 6 1; ni 2 positive integer; 1 6 i 6 m;
where ri and ni are the reliability and the number of components in the ith subsystem respectively; f(Æ) is the
objective function for the overall system reliability; g(Æ) is the constraint function and l is the resource limita-
tion; m is number of subsystems. The goal is to determine the number of component and the components’
reliability in each subsystem so as to maximize the overall system reliability. This problem belongs to the cat-
egory of constrained nonlinear mixed-integer optimization problems.
1558 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567

For solving such kind of mixed-integer reliability problems, most efforts have been devoted to nonlinearly-
constrained reliability-redundancy allocation problems. It’s known that the nonlinear mixed-integer program-
ming problems are more difficult than pure redundancy allocation problems. Tillman et al. [24] and Gopal
et al. [7] applied heuristic approaches to solve the reliability-redundancy allocation problem. Kuo et al. [18]
demonstrate the Lagrange multipliers with branch-and-bound method for a series system with five subsystems.
The interative heuristic method and surrogate dual approach are used to solve this mixed-integer reliability
design problems by Xu et al. [26] and, Hikita et al. [9], respectively. However, most of those require derivatives
for all nonlinear constraint functions. That makes the exact optimal solutions to the reliability-redundancy
allocation problem be not derived easily because of the highly computational complexity. For overcome this
difficulty, Yokota et al. [27] and Hsieh et al. [13] applied genetic algorithms to solve these mixed-integer reli-
ability optimization problems. As they have reported, the solutions obtained by GAs can solve these problems
efficiently and effectively.
Recently, artificial immune algorithms (IAs), which were originally proposed by Jerne [15], have been
widely studied and applied to a variety of engineering optimization problems, usually of a combinatorial nat-
ure. Owing to numerous reports of successful applications of these innovative algorithms, IAs have attracted
more recent attention than most other heuristic/metaheuristic methods in various field of optimization, includ-
ing the reliability problems. For instance, Chen and You [1] used IAs to solve redundancy allocation problems
(integer reliability problems) and reported effective solutions. In our knowledge, it may be the first trial of
applying IAs to mixed-integer reliability optimization problems in the literature.
The metaheuristic optimization approach employing penalty guided based IAs for the reliability-redun-
dancy allocation problems is proposed in our work. The merits of IAs lie in pattern recognition, memorization
capabilities [14] and the theory was originally proposed by Jerne [15]. Compared with other metaheuristic
approaches such as genetic algorithms and evolution strategies, the IAs base approach has very distinct char-
acteristics: (i) the diversity is embedded by calculating the affinity (ii) the self-adjustment of the immune
response is accomplished by the boost or restriction of antibody generations. These characteristics are also
the advantages for solving the mixed-integer programming problems because: (i) the diversities of the feasible
spaces can be better ensured, i.e., the global optimum can be more likely approached, and (ii) a population of
antibodies in IAs can operate simultaneously so that the possibility of paralysis in the whole process can be
reduced.
This paper is arranged as follows: in the next section the proposed penalty guided IAs approach is intro-
duced. Section 3 provides a description of the four test problems and the related data used for this study. The
results of the experiments are also discussed. Finally, the conclusion of the paper is summarized and the direc-
tions for future research are described.

2. Immune algorithm based approach

The natural immune system is a very complicated system for defense against pathogenic organisms. A two-
tier line of defense is in the immune system comprehending the innate immune system and the adaptive
immune system. The fundamental components are lymphocytes and antibodies [6]. The cells of the innate
immune system are immediately available to combat against a wide variety of antigen without previous expo-
sure to them. The antibody production in response to a determined infectious agent (antigen) is the adaptive
immune response mediated by lymphocytes which are responsible for recognition and elimination of the path-
ogenic agents [3]. The cells in the adaptive system are able to develop an immune memory so that they can
recognize the same antigenic stimulus when it is presented to the organism again. Also, all the antibodies
are produced only in response to specific infections. B-cells (B-lymphocytes) and C-cells (T-lymphocytes)
are two major types of lymphocytes. They carry surface receptor molecules capable of recognizing antigens.
The B-cells produced by the bone marrow show a distinct chemical structure and can be programmed to make
only one antibody that is placed on the outer surface of the lymphocyte to act as a receptor. The antigens will
only bind to these receptors with which it makes a good fit [10].
The major task of the immune system is to distinguish and eliminate the intruders of the organism so that
the immune system must has the capability of self/non-self discrimination. As mentioned previously, various
antibodies can be produced and then can recognize the specific antigens. The portion of antigen recognized by
 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567 1559

antibody is called epitope which acts as an antigen determinant. Every type of antibody has its own specific
antigen determinant which is called idiotope. Moreover, in order to produce enough specific effector cells to
against an infection, and activated lymphocyte has to proliferate and then differentiate into these effector cells.
This process is called clonal selection [25] and followed by the genetic operations such that a large clone of
plasma cell is formed. Therefore, the antibodies can be secreted and ready to bind antigens. According to
above facts, Jerne [14] proposed an idiotype network hypothesis which is based on the clonal selection theory.
In his hypothesis, some types of recognizing sets are activated by some antigens and produce an antibody
which will then activate other types of recognizing sets. Through the way, the activation is propagated through
entire network of recognizing sets via antigen-antibody reactions. It is noted that the antigen identification is
not done by a single or multiple recognizing sets but by antigen-antibody interactions. The more details are
referred to Huang [10,11]. From this point of view, for solving the reliability-redundancy allocation optimi-
zation problems, the antibody and antigen can be looked as the solution and objection function respectively.

2.1. Computation procedures

The computation procedures of the proposed immune algorithms based approach illustrated in Fig. 1 work
as follows and the discussion comes in sequence:

Step 1. Randomly generate an initial population of antibodies.

Step 2. Evaluate the corresponding affinity value for each individual.
Step 3. Choose the best k individual with highest affinity values and then clone these k antibodies.
Step 4. The set of the clones in previous step will suffer the genetic operation process, i.e., crossover and muta-
tion [19].
Step 5. Update the memory set. The process includes replacement and elimination.
• Firstly, calculate the new affinity values of these new individuals (antibodies). Select those individ-
uals who are superior to the individuals in the memory set, and then the superior individuals replace
the inferior individuals in the memory set.
• Secondly, antibodies will be eliminated while their structures are too similar so that the individuals in
the memory set are able to keep the diversity.

Generate initial Calculate the

(Step 1) affinity values (Step 2)
population

(Step 6) Chooseand clone the

(Step 7)
No best k antibodies
Yes
Best solution Stop? (Step 3)
Temporary population
of antibodies

Memory Set Adaptive Operation

• Crossover
• Mutation
(Step 4)
(Step 5)
Updating operation
• Replacement Population of the
• Elimination Adapted antibodies

Fig. 1. The immune based approach.

1560 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567

Step 6. Check the stopping criterion, if not stop then go to Step 2. Otherwise go to next step.
Step 7. Stop. The optimal or near optimal solution(s) can be provided from the memory set.

Because of the soul of diversity in the IAs, the quality of solutions in the feasible space can be better guar-
anteed and obtained. So, a suppression process (diversity embodiment) is needed and shown on the Step 5 in
the proposed IAs procedure. The diversity in each pair of antibody i (Abi) and antibody j (Abj) can be eval-
uated by calculating their affinity (fij) by following way:
fij ¼ kAbi  Abj k; for all i and j:
While the affinity between each pair of antibodies in memory is obtained, the antibodies will be eliminated if
the affinity is less than the predefined threshold. So, the diversity of the antibodies in memory is embodied. It is
noted that the way of evaluating affinities of Ab–Ab and Ab–Ag are distinct. The procedure of evaluating the
antibodies is to calculate the Ab–Ag affinity for each antibody that will be illustrated in the Section 2.3.
In our implementation, the integer solutions are represented by strings of binary digits. Each string consist-
ing of substring includes the type of component and redundant levels for each subsystem. The details have
been described in next section. In the above procedures, the clonal selection and affinity maturation processes
are described in details by De Castro and Von Zuben [4]. The stopping criterion is the maximum iterations in
this article.

2.2. The representation mechanism and embodiment of diversity

The solution representation for IAs can be used in the same manner to that of genetic algorithms. In our
implementation, the antibody (solution) is represented by a binary string, each string consisting of a substring
for each subsystem. Each subsystem in turn consists of a binary substring representing the level of redundancy
(integer number) and the corresponding reliability of redundant components (real number) type of compo-
nent. An integer number can be obtained by rounding a real number which is represented by a binary string.
It is illustrated in Fig. 2.

2.3. Constrained optimization

For breeding the superior antibodies, to evaluate the antibody is a necessary process for the immune algo-
rithms. The goal of the algorithm is to adapt the unfeasible antibodies to the feasible antigen(s), so as to reduce
the constraint violations of the search for obtaining the optimal or near optimal solutions. Like the majority of
genetic algorithms applications, for handling these constraint violations the penalty function has been defined.
The penalty function increases the penalty for infeasible solutions based on the distance away from the feasible
region. According to the Eq. (2) in the problem formulation, the function has been defined and described as
follows:

SYSTEM
Subsystem 1 Subsystem 2 Subsystem m

Level of redundancy Reliability of each component

1011…01 1010…11
Integer string: n1 Real number string: n2

Fig. 2. Binary string solution representation.

 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567 1561

Assume the individual j within the memory of N. For each individual antibody, the constraint violation
value for the jth individual is defined as,

gðr; nÞ  b; if gðr; nÞ > b;
Vj ¼
0; Otherwise:

Note the objective and solution are deemed as the antigen and antibody respectively. After defining the pen-
alty function, the fitness of each antibody to the antigen (objective) can be obtained. In other words, the affin-
ity between each antibody and antigen is able to be determined. The affinity function (fitness function) of any
Ab to Ag is described below:
RS
Affinity ¼ Pk ;
1þ j¼1 V j

where, k is the number of constraints in the problem. The above affinity value is to be maximized when the
penalty is minimized.

2.4. Genetic operation process

The implementation of genetic operations is the same as in genetic algorithms. It includes the crossover
operator and mutation operator requires the selection of the crossover point(s) and mutation point(s) for each
antibody (string) under a predetermined crossover probability and mutation probability. The crossover oper-
ator provides a thorough search of the sample space to produce good solutions. The mutation operator per-
forms random perturbations to selected solutions to avoid the local optimum. Note the mutation rate must be
small enough to avoid degrading the performance.

3. Numerical examples and discussion

To evaluate the performance of the proposed approach for the mixed-integer nonlinear reliability design
problems, four test problems (P1  P4) are solved. The nonlinear constraints of the first three examples used
by Hikita et al. [9], Kuo et al. [17] and Xu et al. [26] are also used in the three examples which include a series
system, series-parallel system and complex (bridge) system. The fourth example is an overspeed protection sys-
tem which was investigated by Yokota et al. [27] and Dhingra [5]. All the above problems of maximizing the
systems reliability subject to multiple nonlinear constraints can be stated as the mixed-integer nonlinear pro-
gramming problem. For each problem, both the component reliabilities and redundancy allocations are to be
decided simultaneously.
The four reliability-redundancy problems are formulated below.

P1. Series system (Fig. 3 [9])

Y
m
Max f ðr; nÞ ¼ Ri ðni Þ
i¼1

X
m
Subject to g1 ðr; nÞ ¼ wi v2i n2i 6 V ;
i¼1

X
m
b
g2 ðr; nÞ ¼ ai ð1000= ln ri Þ i ðni þ expðni =4ÞÞ 6 C;
i¼1

X
m
g3 ðr; nÞ ¼ wi ni expðni =4Þ 6 W ;
i¼1

0 6 ri 6 1; ni 2 positive integer; 1 6 i 6 m
1562 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567

1 2 3 4 5

Fig. 3. Series system.

1 2

3
5
4

Fig. 4. Series-parallel system.

1 2

3 4

Fig. 5. Complex (bridge) system.

P2. Series-parallel system (Fig. 4 [9])

Max f ðr; nÞ ¼ 1  ð1  R1 R2 Þð1  ð1  R3 Þð1  R4 ÞR5 Þ
Subject to g1 ðr; nÞ; g2 ðr; nÞ and g3 ðr; nÞ

0 6 ri 6 1; ni 2 positive integer; 1 6 i 6 m
P3. Complex (bridge) system (Fig. 5 [9])
Max f ðr; nÞ ¼ R1 R2 þ R3 R4 þ R1 R4 R5
þ R2 R3 R5  R1 R 2 R3 R4  R1 R2 R 3 R5  R1 R2 R4 R5
 R1 R3 R4 R5  R2 R3 R4 R5 þ 2R1 R2 R3 R4 R5
Subject to g1 ðr; nÞ; g2 ðr; nÞ and g3 ðr; nÞ
0 6 ri 6 1; ni 2 positive integer; 1 6 i 6 m
P4. Overspeed protection system [5,27].
The fourth example is an overspeed protection system of gas turbine with a time related cost function
[5,27]. It is assumed that during the operating time, the system components must not fail. This reliabil-
ity-redundancy allocation problem is formulated as the following mixed-integer nonlinear programming
problem.
 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567 1563

Y
m
Max f ðr; nÞ ¼ ½1  ð1  ri Þni 
i¼1
Xm
Subject to h1 ðr; nÞ ¼ vi n2i 6 V
i¼1
X
m
h2 ðr; nÞ ¼ Cðri Þ  ½ni þ expðni =4Þ 6 C
i¼1
Xm
h3 ðr; nÞ ¼ wi ni expðni =4Þ 6 W
i¼1

1:0 6 ni 6 10; ni 2 positive integer;

6
0:5 6 ri 6 1  10 ; rj 2 real number;
1:0 6 ni 6 10; ni 2 positive integer;
6
0:5 6 ri 6 1  10 ; ri 2 real number;
bi
where, Cðri Þ ¼ ai ½T = lnðri Þ is the cost of each component with reliability ri at subsystem i; T is the operating
time during which the component must not fail.

The input parameters defining the specific instances of the four problems have the same values as Kuo et al.
[17], Xu et al. [26], Hikita et al. [9], Hsieh et al. [13], Dhingra [5], and Yokota et al. [27], and are show in Tables
1–3.
Our artificial immune algorithm is implemented in MATLAB on the Pentium-4 3.2 GHz PC with the fol-
lowing parameters: memory size = 90, mutation rate = 0.01, crossover rate = 0.92 and the maximum clone
number = 15. Then number of generations was specified to be 150. The determination of immune algorithm’s

Table 1
Data used in series system (P1) and complex system (P3) [9]
i 105 ai bi wi v2i wi V C W
1 2.33 1.5 1 7 110 175 200
2 1.450 1.5 2 8
3 0.541 1.5 3 8
4 8.050 1.5 4 6
5 1.950 1.5 2 9

Table 2
Data used in series-parallel system (P2) [9]
i 105 ai bi wi v2i wi V C W
1 2.500 1.5 2 3.5 180 175 100
2 1.450 1.5 4 4.0
3 0.541 1.5 5 4.0
4 0.541 1.5 8 3.5
5 2.100 1.5 4 4.5

Table 3
Data used in overspeed protection system (P4) [5,27]
Stage 105 aj bj vi wi V C W T
1 1 1.5 1 6 250.0 400.0 500.0 1000 h
2 2.3 1.5 2 6
3 0.3 1.5 3 8
4 2.3 1.5 2 7
1564 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567

parameters is a significant problem for the immune algorithm implementation. However, there is no formal
methodology to solve the problem because different value-combinations of the parameters result to different
characteristics as well as different performance of immune algorithms. Therefore, one should note that the best
values for the artificial immune algorithm parameters are case-dependent and based upon the experience from
preliminary runs. The numerical results are shown in Table 4 through 7, in which the best solutions of each
problem are reported and compared with solutions reported previously in the literature.
Table 4 indicate that the solution of the series problem found by our proposed approach is just slightly bet-
ter than the solution found by Xu et al. [26], but is comparatively better than the solutions found by Hikita
et al. [9], Kuo et al. [17], and Hsieh et al. [13]. Table 5 shows that the solution of series-parallel problem found
by the proposed approach is better than those reported by Hikita et al. [9] and Hsieh et al. [13]. Compared
with the solutions found by Dhingra [5] and Yokota et al. [27] in the Table 6, the solutions found by proposed
method are relatively with more significant improvement. In Table 7, the solution found by the proposed
approach is much better than the previous best know solution by Yokota et al. [27] and Dhingra [5].
Although the proposed penalty guided immune algorithm found better solutions for all the four example
problems, some of the improvement looks extremely tiny, for instance: in problems P1, P2 and P3 where some
differences are on the order of 106 or 105. Especially, the solution comparison between the proposed method
and genetic approach for the second and third problems shows that the improvement seems to be relatively
smaller (see Tables 5 and 6). However, the tiny improvements in reliability are often hard to be achieved in

Table 4
Comparison of the best proposed approach solutions with other algorithms for series system (P1)
Hikita et al. [9] Kuo et al. [17] Xu et al. [26] Hsieh et al. [13] Chen
n (3, 2, 2, 3, 3) (3, 3, 2, 3, 2) (3, 2, 2, 3, 3) (3, 2, 2, 3, 3) (3, 2, 2, 3, 3)
0.777143 0.77960 0.77939 0.779427 0.779266
0.867514 0.80065 0.87183 0.869482 0.872513
r 0.896696 0.90227 0.90288 0.902674 0.902634
0.717739 0.71044 0.71139 0.714038 0.710648
0.793889 0.85947 0.78779 0.786896 0.788406
Rs 0.931363 0.92975 0.931677 0.931578 0.931678
MPI (%) 0.459% 2.744% 0.001% 0.146%
Slacks of (g1  g3) 27 27 27 27 27
0.000000 0.000010 0.013773 0.121454 0.001559
7.518918 10.57248 7.518918 7.518918 7.518918
Note: Slack is the unused resources.
MPI (%) = (Rs_Chen  Rs_other)/(1  Rs_other).

Table 5
Comparison of the best proposed approach solutions with other algorithms for series-parallel system (P2)
Hikita et al. [9] Hsieh et al. [13] Chen
n (3, 3, 1, 2, 3) (2, 2, 2, 2, 4) (2, 2, 2, 2, 4)
0.838193 0.785452 0.812485
0.855065 0.842998 0.843155
r 0.878859 0.885333 0.897385
0.911402 0.917958 0.894516
0.850355 0.870318 0.870590
Rs 0.99996875 0.99997418 0.99997658
MPI (%) 30.325% 10.248%
Slacks of (g1  g3) 53 40 40
0.000000 1.194440 0.002627
7.110849 1.609289 1.609289
Note: Slack is the unused resources.
MPI (%) = (Rs_Chen  Rs_other)/(1  Rs_other).
 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567 1565

Table 6
Comparison of the best proposed approach solutions with other algorithms for complex system (P3)
Hikita et al. [9] Hsieh et al. [13] Chen
n (3, 3, 2, 3, 2) (3, 3, 3, 3, 1) (3, 3, 3, 3, 1)
0.814483 0.814090 0.812485
0.821383 0.864614 0.867661
r 0.896151 0.890291 0.861221
0.713091 0.701190 0.713852
0.814091 0.734731 0.756699
Rs 0.99978937 0.99987916 0.99988921
MPI(%) 47.401% 8.317%
Slacks of (g1  g3) 18 18 18
1.854075 0.376347 0.001494
4.264770 4.264770 4.264770
Note: Slack is the unused resources.
MPI (%) = (Rs_Chen  Rs_other)/(1  Rs_other).

Table 7
Comparison of the best proposed approach solutions with other algorithms for complex system (P4)
Dhingra [5] Yokota et al. [27] Chen
k (6, 6, 3, 5) (3, 6, 3, 5) (5, 5, 5, 5)
0.81604 0.965593 0.903800
0.80309 0.760592 0.874992
r 0.98364 0.972646 0.919898
0.80373 0.804660 0.890609
0.99961 0.999468 0.999942
MPI (%) 85.128% 89.098%
Slacks of (h1  h2) 65 92 50
0.064 70.733576 0.002152
4.348 127.583189 28.803701
Note: Slack is the unused resources.
MPI (%) = (Rs_Chen  Rs_other)/(1  Rs_other).

high reliability applications. For measuring the improvement, MPI (maximum possible improvement) can be
sued to measure the amount improvement of the solutions found by the proposed approach to the previous
best know solutions. MPI is the fraction that the best feasible solution achieved of the maximum possible
improvement, considering that reliability 61 [2] and it is described as:
MPI(%) = (Rs_Chen  Rs_other)/(1  Rs_other), where Rs_Chen represents the system reliability obtained by the
proposed approach and Rs_other represents the system reliability obtained by other approaches in literature.
By using the index, it shows the proposed approach made larger improvements in P2  P4. From Tables 5–
7, it can be seen that the minimum improvement for P2, P3 and P4 are 10.248%, 8.317% and 85.128% respec-
tively. Especially in Table 7 (P4), the solution found by the proposed approach is much superior to those by
Dhingra [5], and Yokota et al. [27]. Again, it has to be emphasized that even very small improvements in reli-
ability are often hard to be obtained in high reliability applications. Moreover, the solutions found by the pro-
posed approach can dominate any other methods for the four example problems discussed in literature. In
other words, then, one could say that proposed penalty guided immune algorithm is able to find solutions
of quality comparable to those previously published in the literature.
It looks that GAs and IAs are similar, but there are an essential difference in the memory adopting system
and the production system of various antibodies. It allows the global optimum to be acquired by using IAs
form many optimization problems. The main cause is that the diversity characteristic in memory of IAs makes
the proposed approach with more chances explore the global optimal solution. However, this merit of the IAs
may become its disadvantage while the CPU time is taken into account. Although more CPU time is taken in
1566 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567

IAs than in GAs, it is still worth to do so since obtaining a system design with higher reliability is very difficult
and important in the real world applications.
According to the above observations, we concluded that the performance of the proposed approach is supe-
rior to the other methods when used to find the maximum reliability for these nonlinearly mixed-integer reli-
ability design problems (CPU time is ignored). Again, in these optimization problems, both the redundancy
(number of redundant components) and the corresponding reliability of each component in each subsystem
under multiple constraints are to be decided simultaneously.

4. Conclusions

This paper presents a penalty guided immune algorithm for solving various reliability design problem,
which include series systems, series-parallel system, complex (bridge) system and overspeed protection system.
Unlike the traditional GAs based approaches, IAs based approach preserves diversity in the memory so that it
is able to discover new optima over time. Therefore, the convergence of IAs based approach is never com-
pleted and this diversity acts like a preventive measure. This notion of viability of enabling further adaptations
is precisely what GAs were lacking and this may become the cause why IAs based approach provides superior
solution than GAs based approaches do. As demonstrated in the previous section, the best solutions found by
our IAs are all better than or tie the well-know best solutions by other heuristic methods for mixed-integer
reliability problems. While the improvement may be so small as not to be significant, our limited experience
suggests that the IAs find solutions which are of a quality and are comparable to that of other heuristic algo-
rithms while the CPU time is ignored. The proposed method achieves the global solution or finds a near-global
solution for each example problem tested.

Acknowledgement

The research is supported by grants from National Science Council, Taiwan under contract NSC 93-2213-
E-150-012.

References

[1] T.-C. Chen, P.-S. You, Immune algorithm based approach for redundant reliability problems, Computers in Industry 56 (2005) 195–
205.
[2] D.W. Coit, A.E. Smith, Reliability optimization of series-parallel systems using a genetic algorithm, IEEE Transaction on Reliability
45 (1996) 254–260.
[3] L.N. De Castro, J. Timmis, Artificial Immune Systems: A new Computational Intelligence Approach, Springer, New York, 2002.
[4] L.N. De Castro, F.J. Von Zuben. The clonal selection algorithm with engineering applications, in: Workshop Proceedings of the
GECCO 2000, 2000, 36–37.
[5] A.K. Dhingra, Optimal apportionment of reliability & redundancy in series systems under multiple objectives, IEEE Transactions on
Reliability 41 (4) (1992) 576–582.
[6] J.D. Farmer, N.H. Packard, A.S. Perelson, The immune system, adaptation, and machine learning, Physica 22D (1986) 187–204.
[7] K. Gopal, K.K. Aggarwal, A new method for solving reliability optimization problem, IEEE Transaction on Reliability R 28 (1978)
36–38.
[8] F.S. Hillier, G.J. Lieberman, Operations Research, McGraw-Hill, New York, 1995.
[9] M. Hikita, Y. Nakagawa, H. Harihisa, Reliability optimization of systems by a surrogate constraints algorithm, IEEE Transactions
on Reliability R 41 (3) (1992) 473–480.
[10] S.J. Huang, Enhancement of thermal unit commitment using immune algorithms based optimization approaches, Electrical Power
and Energy Systems 21 (1999) 245–252.
[11] S.J. Huang, An immune-based optimization method to capacitor placement in a radial distribution system, IEEE Transaction on
Power Delivery 15 (2) (2000) 744–749.
[12] Y.-C. Hsieh, A linear approximation for redundant reliability problems with multiple component choices, Computers and Industrial
Engineering 44 (2003) 91–103.
[13] Y.-C. Hsieh, T.-C. Chen, D.L. Bricker, Genetic algorithms for reliability design problems, Microelectronic Reliability 38 (1998) 1599–
1605.
[14] N.K. Jerne, The immune system, Scientific America 229 (1) (1973) 52–60.
[15] N.K. Jerne, Clonal selection in lymphocyte network, in: G.M. Edelman (Ed.), Cellular Selection and Regulation in the Immune
Response, Raven Press, New York, 1974.
 T.-C. Chen / Applied Mathematics and Computation 182 (2006) 1556–1567 1567

[16] W. Kuo, V.R. Prasad, An annotated overview of system-reliability optimization, IEEE Transaction on Reliability 49 (2) (2000) 176–
187.
[17] W. Kuo, C.L. Hwang, F.A. Tillman, A note on heuristic methods in optimal system reliability, IEEE Transactions on Reliability R 27
(1978) 320–324.
[18] W. Kuo, H. Lin, A. Xu, W. Zhang, Reliability optimization with the Lagrange multiplier and branch and bound technique, IEEE
Transaction on Reliability R 36 (1987) 624–630.
[19] Z. Michalewicz, Genetic Algorithm + Data structures = Evolution Programs, Springer-Verlag, Berlin Heidelberg, New York, 1994.
[20] K.B. Misra, A simple approach for constrained redundancy optimization problem, IEEE Transactions on Reliability R 21 (1972) 30–
34.
[21] K.B. Misra, J. Sharma, A new geometric programming formulation for a reliability problem, International Journal of Control 18
(1973) 497–503.
[22] Y. Nakagawa, S. Miyazaki, Surrogate constraints algorithm for reliability optimization problems with two constraints, IEEE
Transaction on Reliability R 30 (1981) 175–180.
[23] K.S. Park, Fuzzy apportionment of system reliability, IEEE Transactions on Reliability R 36 (1987) 129–132.
[24] F.A. Tillman, C.L. Hwang, W. Kuo, Determining component reliability and redundancy for optimum system reliability, IEEE
Transaction on Reliability R 26 (3) (1977) 162–165.
[25] I.L. Weissman, M.D. Cooper, How the immune system develops, Scientific American 269 (3) (1993) 33–40.
[26] Z. Xu, W. Kuo, H.H. Lin, Optimization limits in improving system reliability, IEEE Transactions on Reliability R 39 (1990) 51–60.
[27] T. Yokota, M. Gen, Y.-X. Li, Genetic algorithm for nonlinear mixed-integer programming problems and its application, Computers
and Industrial Engineering 30 (4) (1996) 905–917.