Reliability Engineering and System Safety 65 (1999) 229–238

Application of genetic algorithm for reliability allocation in nuclear

power plants
Joon-Eon Yang*, Mee-Jung Hwang, Tae-Yong Sung, Youngho Jin
Integrated Safety Assessment Team, Advanced Nuclear Technology Development, Korea Atomic Energy Research Institute, P.O. Box 105, Yusong-Gu,
Taejon 305-600, South Korea
Received 1 June 1998; received in revised form 8 December 1998; accepted 8 December 1998

Abstract
Reliability allocation is an optimization process of minimizing the total plant costs subject to the overall plant safety goal constraints.
Reliability allocation was applied to determine the reliability characteristics of reactor systems, subsystems, major components and plant
procedures that are consistent with a set of top-level performance goals; the core melt frequency, acute fatalities and latent fatalities.
Reliability allocation can be performed to improve the design, operation and safety of new and/or existing nuclear power plants. Reliability
allocation is a kind of a difficult multi-objective optimization problem as well as a global optimization problem. The genetic algorithm,
known as one of the most powerful tools for most optimization problems, is applied to the reliability allocation problem of a typical
pressurized water reactor in this article. One of the main problems of reliability allocation is defining realistic objective functions. Hence, in
order to optimize the reliability of the system, the cost for improving and/or degrading the reliability of the system should be included in the
reliability allocation process. We used techniques derived from the value impact analysis to define the realistic objective function in this
article. 䉷 1999 Elsevier Science Ltd. All rights reserved.
Keywords: Reliability allocation; Genetic algorithm; Value impact analysis

Nomenclature RT Reactor Trip System

SIT Safety Injection Tank
AFWS Auxiliary Feedwater System
SWS Service Water System
BD Bleed System
SLOCA Small LOCA
CDF Core Damage Frequency
SR Steam Removal System
DG Diesel Generator
VIA Value Impact Analysis
EPS Electric Power Supply System
HPSI High Pressure Safety Injection System
IA Instrument Air System
LOCA Loss of Coolant Accident
LLOCA Large LOCA 1. Introduction
LCD Late Cancer Death
LCF Latent Cancer Fatality Protecting the public from the risk of NPPs by increasing
LOFW Loss of Main Feedwater the safety of NPPs is the main goal of the nuclear safety
LOOP Loss of Offsite Power study. To accomplish this goal, various approaches are
LPSI Low Pressure Safety Injection System adopted in several aspects: (1) the diversity and redundancy
MFWS Main Feedwater System concepts in the design of NPPs, (2) the surveillance and
MLOCA Medium LOCA testing of components, and (3) the various safety analyses
NPPs Nuclear Power Plants such as PSA. Our ultimate goal is to keep the risk from NPPs
PDS Plant Damage State as zero, however, it is impossible to achieve this goal in the
PWR Pressurized Water Reactor real world owing to various technical and economical
PSA Probabilistic Safety Assessment restrictions. So an approach called reliability allocation
emerged [1–7]. Reliability allocation is a kind of optimiza-
* Corresponding author. Tel.: 082-42-868-8640; Fax: 082-42-868-8374. tion problem for minimizing the total plant costs subject to
E-mail address: jeyang@nanum.kaeri.re.kr (J.-E. Yang) the overall plant safety goal constraints [1].
0951-8320/99/$ - see front matter 䉷 1999 Elsevier Science Ltd. All rights reserved.
PII: S0951-832 0(98)00103-3
230 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238

Fig. 1. Flow chart of genetic algorithm.

Reliability allocation was applied to determine the relia-
bility characteristics of reactor systems, subsystems, major
components and plant procedures that are consistent with a
set of top-level performance goals: the core melt frequency,
acute fatalities and latent fatalities. Reliability allocation
can be performed to improve the design, operation and
safety of a new and/or existing NPP. The importance of
reliability allocation is emphasized in several aspects: (1)
to minimize the risk of nuclear power plants under various
constraints, (2) to allocate limited resources effectively, and
(3) to reduce over-design.
Until now, various approaches were proposed to solve
this problem. However, most approaches have some limita-
tions in satisfying all optimization objectives [1]. Most opti-
mization techniques used in the previous approaches are
based on conventional optimization techniques that can
not guarantee the global optimum for nonlinear multi-
objective problems like reliability allocation. Hence the
previous approaches used some assumptions to simplify
the problems. Recently, it has become known that the
genetic algorithm is one of the most powerful tools for
many optimization problems [8]. The genetic algorithm
can provide a global optimal solution with the proper choice
of related parameters. It also enables us to provide various
constraints and to define complex objective functions. It was
applied to the optimization of reliability and test interval
problems [9–11]. Hence, in this article, the genetic algo-
rithm is applied to the reliability allocation process of NPPs.
One of the main problems of reliability allocation is
defining the realistic objective functions. For simplicity,
 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238 231

Table 1
Data to quantify the health and investment risks

Plant damage state Health risk Investment risk Total risk

1 Severe core damage or core 5.0 × 10 person rem/

4
$1162–$3136 million $1167–3,186 million
melt; significant radioisotope event, ( ˆ $5–$50 median: $2149 million median: $2176.5 million
release to containment million/event)
2 Small LOCA leading to 3.8 × 10 4 person rem/ $329–$924 million $332.8–962 millions
containment cleanup, valve and event, ( ˆ $3.8–$38 median: $626.5 millions median: $647.4 million
vessel repair to containment million/event)
3 Possible damage to steam $32–$243 millions $32–243 millions
generator; minor containment median: $137.5 million median: $137.5 million
cleanup and equipment checkout
4 Possible primary system water $1–$6 millions median: $1–6 millions median:
loss; little or no spill into $3.5 million $3.5 million
containment; no core or
equipment damage

until now, simplified objective functions were used in most
reliability allocation methods. In some systems, it is very
difficult to increase the reliability of the systems; however, it
is easy to improve the reliability of some systems with very
little cost. So in order to optimize the reliability of a system,
the cost for increasing and/or decreasing the reliabilities of
the system and/or the CDF should be incorporated into the
reliability allocation process. To define the realistic objec-
tive function, we used techniques derived from the VIA
[12–14]. The VIA can provide the best alternative design
based on the given criteria. In this article, VIA is used as a
tool to define the objective function of the genetic algo-
rithm. The value and impact for increasing and/or decreas-
ing the unavailability of systems are estimated based on
several aspects such as costs for design and hardware
changes, availability, licensing and etc. The costs and bene-
fits related to the changes of total CDF are also estimated.
In addition, various constraints owing to the realistic
situation such as the requirements from the regulatory
body are also provided as additional constraints, e.g., the

Fig. 2. Example of event tree: Small LOCA.

232 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238
reliability target of AFWS is 1.0 × 10 ⫺3/yr in PWR. The
acceptable ranges of each system’s unavailability are also
defined as the constraints.
2. Genetic algorithm
A genetic algorithm mimics the real physical world. A
genetic algorithm generates the initial population of solu-
tions. This population evolves over successive generations
based on the survival of fitness. The operations such as
reproduction, cross over and mutation are performed on
the populations and the fitness of each individual is evalu-
ated. Based on the new fitness of each individual, the popu-
lation of next generation is produced probabilistically. As
time moves on, the individuals with poor fitness will disap-
pear, and the individuals with high fitness will survive. The
genetic algorithm can search huge space rapidly based on
the implicit parallelism. The flowchart of the genetic algo-
rithm is shown in Fig. 1 [8].
Genetic algorithms are appropriate for various optimiza-
tion problems including multi-modal and multi-objective
problems. For more details of genetic algorithms, refer to
Ref. [8]. This genetic algorithm is applied to the reliability
allocation process of NPPs in this article.
The problem specific features of a genetic algorithm such
as objective functions and selection of parameters will be
explained in later sections.
3. Problems to be solved
Reliability allocation is based on a PSA model. Thus a
simplified PSA model is developed for a typical PWR. We
developed simplified event trees for the following four
major initiating events.
1. Loss of coolant accident group
• large and medium LOCA: the frequency is assumed as
6.60 × ⫺6/yr;
• small LOCA: the frequency is assumed as 2.43 × ⫺5/
yr.
2. Transient group
• loss of main feedwater: the frequency is assumed as
5.40 × ⫺3/yr;
• loss of offsite power: the frequency is assumed as
6.15 × ⫺4/yr.
As the effect of core damage is different according to the
severity of accident sequences, the core damage state is
divided into four PDSs. The descriptions for each PDS are
given in Table 1. PDS 1–4 are assigned to the end points of
event trees according to the characteristics of each accident
sequence. The accident sequence are quantified for PDSs.
An example of an event tree (i.e., SLOCA event tree) is
shown in Fig. 2.
We also developed simplified fault trees for the following
12 systems:
1. Reactor Trip System (RT),
2. Bleed System (BD),
3. Safety Injection Tank (SIT),
4. High Pressure Safety Injection System (HPSI),
5. Low Pressure Safety Injection System (LPSI),
6. Main Feedwater System (MFWS),
7. Auxiliary Feedwater System (AFWS),
8. Steam Removal System (SR),
9. Electric Power Supply System (EPS),
10. Diesel Generator (DG),
11. Service Water System (SWS), and
12. Instrument Air System (IA).
As the main goal of reliability allocation in this article is to
determine the reliability target in the system level, the
developed fault trees are simple ones that consist of system
failures (hardware failures and related human errors) and
those of related support systems.
The problem to be solved is how to optimize the reliabil-
ities of each system modeled in this simplified PSA model
with the consideration for cost. There are a lot of items that
affect the reliabilities of systems: system designs, compo-
nent hardware failure rate, human error probability, and
maintenance practice. The effects of each item will be
discussed, when we define the objective function in Section
4.
4. Define objective function
One of the main problems in the reliability allocation is
defining the realistic objective functions. Until now, most
reliability allocation methods have used simplified objective
functions for simplicity. In the real world, however, the
costs of increasing and/or decreasing reliabilities of a
system depend upon the design and characteristics of the
system. So in order to optimize the reliability of a system,
the costs of improving and/or degrading the reliability of
that system should be incorporated into the reliability allo-
cation process. For this, we used techniques derived from
the VIA to define realistic objective functions of the genetic
algorithm.
The VIA can provide the best alternative designs based
on the given criteria [12–14]. However, it is difficult to
estimate the effects of combined alternative designs in the
overall aspect by VIA. VIA is a bottom–up approach.
However, the reliability allocation is a top–down approach,
which decides the range of reliabilities for each system in
order to minimize the overall cost. The value and impact of
increasing and/or decreasing the unavailability of systems
are estimated in several aspects: cost for design and hard-
ware changes, availability, licensing and etc. The cost and
benefit related to the changes of PDS frequencies are also
estimated.
 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238 233

Table 2
Estimation of capital cost for each system.
The costs of the upper limit are obtained by multiplying 1.2 to the costs of the base case, and those of the lower limit are obtained by multiplying 0.8 to the
costs of the base case. The systems marked by a are classified as the safety class, hence, the weighting factor 1.2 is multiplied to the summed cost to derive
capital cost

System Tank Pump Valves Base cost Redundancy Capital cost ( × $1000)
($270 000) ($80 000–11 000) ($5000) ($300 000)
Lower limit Base case Upper limit

RS a 1 4 1488 1860 2232

AFW a 3 4 49 1 4 1543 1928 2314
MFW 4 52 1 4 2016 2520 3024
SR 55 1 2 372 465 558
BD a 8 1 2 263 328.8 395
HPSI a 0.5 2 52 1 2 711 889 1067
LPI/SDC a 0.5 2 32 1 2 654 817 981
SIT a 4 0 2 1 4 1283 1603 1924
EPS 1 1 1040 1300 1560
DG 1 2 4000 5000 6000
SWS 2 4 62 1 4 1141 1426 1711
IA 2 2 30 2 2 880 1100 1320

Actually, it is difficult to estimate health effects in mone-
tary units. Such an approach can cause arguments. However,
it is convenient to use a uniform unit to determine the best
choice even though the monetary unit is not the best one. So,
in this article, we use a monetary unit for the objective
function. The monetary data used in this article are not the
exact ones. The data are used only for comparisons among
alternatives. Our goal is to minimize the overall cost while
keeping safety goals such as reliability targets for systems
and/or the limits of PDS.
The attributes of the objective function consist of (1)
health risk, (2) investment risk, (3) licensability, and (4)
financial impact [12]. Among these attributes, we assumed
that the licensability does not cause big differences in the
cost as all systems are to be reviewed by regulatory bodies.
We also assumed that there are no innovative design
changes in the systems.
The attributes for health and investment risk are esti-
mated as shown in the Table 1. The values for each item
are estimated based on the following backgrounds.
1. Health risk:
• LCF dominates costs owing to an accident, so acute
cancer fatality is assumed to be negligible.
• It is estimated that $100–1000/person-rem, 10 000
man-rem/LCD [12].
• In order to estimate the health risks properly, Level 2
PSA model is required, that is, containment event
trees are to be developed and the effect of the source
term is to be evaluated. However, for simplicity, we
used the dose conversion factor as in Ref. [13].
2. Investments risk:
• It is recognized that single point estimates for the
financial impact of a low probability and high cost
consequence accident are totally inadequate. As a
result, lower and upper bound limits were estimated
for each category of accident [12].
• This risk represents the cost for clean up, repair and/or
decomposition according to PDS after regulatory
response.
Financial impact is estimated based on the following
items:
• capital costs (it is estimated from the cost of compo-
nents),
• operation and maintenance costs,
• system capacity costs,
• system energy costs.
The financial impact is dominated by capital costs; it can
be changed according to the class and/or complexity of
systems. So in this article, we only consider Capital Costs
from among the earlier mentioned four items of financial
impact. The estimation on the capital costs of each system is
shown in Table 2.
In the real world, however, an effective maintenance
strategy can reduce risks with only a small incremental
cost. Moreover, most PSAs show that the two types of
human actions, (i.e., pre- and post-accident human actions),
have great effects on the results of PSA. As the effects of
pre-accident human actions on a system’s unavailability are
very system dependent, it is difficult to assess the effects of
pre-accident human actions in the design stage. So, in this
stage, the pre-accident human actions are not assessed. The
maintenance related pre-accident human actions might be
optimized after the optimization of a system’s reliability as
shown in Ref. [11]. For the effects of post-accident human
actions, in this article, we assumed that the changes of
system reliabilities do not affect the post-accident human
error probabilities. So, the human error probabilities of
234 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238
Fig. 3. Performance results with different seeds and/or parameter values.
post-accident human actions are regarded as constants, i.e.,
these are not the variables to be optimized.
Based on these observations, the costs for changing the
reliabilities of systems are estimated using the following
approach [1]:
cos ti ˆ ai ln…1 ⫹ Ri † ⫹ bi ;
where ai, bi are the constants derived from the cost of base
case, ai, bi ⬎ 0 and Ri is a system unavailability for the ith
alternative.
This function satisfies the following conditions [1]:
1. the cost of a low reliability system is very low,
2. the cost of a high reliability system is very high,
3. the cost is a monotonic increasing function of reliability,
4. the derivative of cost with respect to reliability is a mono-
tonic increasing function of reliability.
From these conditions, the final objective function of the
reliability allocation process is defined as:
X
OBFj ˆ PDSij ‰HRCij ⫹ IRij Š ⫹ FIj ;
i a system and the frequency of an initiating event caused by
where OBFj is the value of objective function probability for
the jth reliability allocation, PDSij, the probability of the ith
PDS based on the jth reliability allocation, HRCij, the value/
impact owing to health risk for ith PDS based on the jth
reliability allocation, IRij, the value/impact owing to
investment risk for ith PDS based on the jth reliability
allocation, and FIj, the financial impact (capital/economical
risk) of the jth reliability allocation.
In addition, the various constraints of the real world, such
as the requirements from the regulatory body, are also
provided as additional constraints, e.g., the reliability target
of AFWS is assumed as 1.0 × ⫺3 in this article. The accep-
table ranges of each system’s unavailability are also defined
as the constraints.
There is one more item we have to consider, i.e., the plant
availability. The change of systems’ reliability causes the
change of initiating event frequencies. Therefore, this aspect
should be included in the objective function. The final
objective function can be represented as follows:
X
OBFj ˆ PDSij …IEijk †‰HRCij ⫹ IRij Š ⫹ FIj ;
i
where IEijk is the change of the kth initiating event frequency
for the ith PDS based on the jth reliability allocation.
It is difficult to derive a relation between the reliability of
the failure of that system as it depends upon the character-
istics of NPPs. In most cases, the front line systems modeled
in PSA do not affect the frequencies of initiating events
greatly. The supporting systems, however, will affect the
frequencies of initiating events greatly. Thus, in this article,
 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238
Fig. 4. Changes of fitness function over generations.
it is assumed that the frequencies of each initiating event are
not affected by the change of system reliability as four initi-
ating events are not related to the supporting systems.
However, as an example, a sample case is analyzed that
incorporates the change of an initiating event frequency
according to the reliability of the system. As only the
frequency of LOFW is affected by the reliability of a system
among four initiating events, we assume that the frequency
of LOFW is changed according to the reliability of the
MFWS.
As mentioned earlier, it is not an easy job to derive the
initiating event frequency from the fault tree. It depends
upon the specific structure of the system and maintenance
practice for that system, and so on. So, based on the follow-
ing considerations, we derive a simple algebraic equation
for the new initiating event frequency of LOFW. In general,
the MFWS consists of running trains and a stand-by train.
Roughly speaking, there is a relation similar as shown:
Q ˆ PR PS ;
where Q is the unavailability of the main feedwater system,
PR, the running failure probabilities of main feedwater oper-
ating trains and PS, the startup failure probability of the
main feedwater stand-by train.
The initiating event frequency can be obtained from the
following formula:
F ˆ g × PS ;
where F is the initiating event frequency of feedwater
system and g, the function of running failure rates. In this
formula, g is a pure initiator and PS is a pure enabler. So we
can derive a simple algebraic equation for the initiating
235
event frequency of the feedwater system, F:
F ˆ g…Q=PR † ˆ …g=PR †Q ˆ hQ;
where h ˆ g=PR . As both g and PR are functions of running
failure rates, we assume that h is almost a constant. In our
example, we derived h from the initial unavailability and
initiating event frequency of MFWS, i.e., 0.1086.
5. Development of problem specific genetic model
The genetic algorithm for this problem is implemented by
using a commercial tool that supports most features of a
genetic algorithm. Based on the developed event trees and
fault trees, the minimal cut sets of each PDS are generated.
The basic events of fault trees, i.e., the failure probabilities
of components, are modeled as chromosomes of each indi-
vidual.
As the genetic algorithm is based on the probabilistic
operations, we perform a kind of sensitivity analyses for
the parameters of genetic algorithm; the different prob-
abilities for operations such as cross over, mutation and etc.
and different seeds for random number generation are tested.
The different population sizes and chromosome lengths are
also tested. Parts of the results are shown in Fig. 3. Based on
the results of such tests, the main parameters used in the
developed model are selected as shown as follows.
• population size: 100,
• length of chromosome: 32 bits,
• cross over rate: 0.9,
• mutation rate: 0.01,
• maximum generation number: 300,
• termination criteria: best fitness unchanged after 50
generations.
236 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238

Fig. 5. Results for unfixed and fixed PDS cases.

The chromosome is composed of 32 genes, which are 6. Results and discussion

implemented in the computer as a binary bit. The genes can
have two values, 0 and 1. So, the chromosome used in this
article can have the precision 1/2 32 as the failure probabilities
of components can have the value in the range from 0 to 1.
The cross over and the mutation operators are applied to a
particular chromosome during a generation as shown in Fig. 1.
In applying the cross over operator, all chromosomes of
two individuals are treated independently, and cross over
operator is applied to each pair of chromosomes.
In the developed model, the constraints limit the values
that the genetic algorithm will search, thus limiting the time
that the genetic algorithm will take to find an optimal solu-
tion. Most examples require a few minutes to converge. The
changes of fitness function over generations for a sample
case are shown in Fig. 4.
The results of reliability allocation are shown in Figs. 5–
7. Fig. 5 shows the changes of system reliabilities for three
typical cases. The base case represents the initial values of
system unavailability before reliability allocation.
The costs owing to health and investment risks have
different characteristics from those owing to financial
impact, that is, the former are indirect costs; and the latter
are direct costs used during the construction of a NPP.
Hence, the reliability allocation is performed for the follow-
ing two cases.
(1) The change of PDS probability is allowed
(‘‘Unfixed’’ in Fig. 5):
This means that the indirect cost is included in the opti-
mization process. Fig. 5 shows that the reliabilities of many

Fig. 6. Results with the different constraints and change of IE frequency.

 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238
Fig. 7. Results with three different cost values.
systems are decreased when compared to the base case.
Such a fact has positive effects in the design and/or opera-
tion of those systems as the designer or operator can have
more flexibility in handling those systems.
(2) The increase of PDS probability is not allowed
(‘‘Fixed’’ in Fig. 5):
This is the case where the increase of PDS probability is
not allowed owing to risk aversion. So, in this case, the
indirect cost is not used for optimization, i.e., the health
and impact risks are not used in the optimization process.
Only the system reliabilities are changed to minimize the
financial impact. The results show that the cost can be
reduced greatly by optimizing the reliabilities of the
systems, while keeping the safety of NPPs at the same level.
The changes in costs are shown in Table 3. In the cost
aspect, the analyzed results show that the total amount of
costs is bounded by the costs related to the system reliabil-
ities as shown in Table 3. As the changes of PDS are small,
the total cost change owing to the PDS change is also small.
In the real world, the change of a systems’ reliability
causes the change of initiating even frequencies. So, in
this article, a sample case is analyzed to show the capability
of the genetic algorithm to handle such a situation. A simple
algebraic equation is used to derive the new frequency of
LOFW based on the new reliability of the main feedwater
system. The results are shown in Fig. 6 as the ‘‘IE-Freq.’’
case. Even though we allow the change of initiating event
Table 3
Changes of costs of three cases
Cost Unfixed ( × $1000) Fixed ( × $1000) IE freq. ( × $1000)
System ⫺8.25 × 10 3 ⫺5.32 × 10 3 ⫺6.80 × 10 3
PDS 1.05 × 10 3 0 ⫺1.70 × 10 2
Total ⫺7.21 × 10 3 ⫺5.32 × 10 3 ⫺6.97 × 10 3
237
frequency, the reliabilities of each system show similar
trends.
The ‘‘Diff. Constr.’’ case of Fig. 6 shows the results of
reliability allocation when an additional constraint is added,
that is, the reliability target of HPSI is given as 1.0 × minus;2.
The unavailability of systems is changed compared to the
reference case, i.e. unfixed PDS case. It seems that the
constraints play very important roles in the optimization
process. So it is important to identify the main constraints
of the real world.
As it is recognized that single point estimates of financial
impact of a low probability and high cost consequence acci-
dent are totally inadequate, the reliability allocations for the
unfixed PDS case are performed again based on the costs of
the lower, upper bound limits and the median value. These
results are shown in Fig. 7. The case using the median value
is the reference case (Unfixed PDS) that is shown in Figs. 5
and 6. The results of Fig. 7 show that the reliabilities of each
system show similar values for all three cases.
7. Conclusions
We implemented the genetic algorithm to the reliability
allocation process of a typical PWR. The simplified PSA
model is developed for a PWR, and VIA is used to define the
objective function. Various cases are analyzed which show
the genetic algorithm is the suitably way to handle the
complex reliability allocation process.
One of the problems we have to pay attention is the
uncertainty that arose from the various sources. That is,
most information used in this article needs further analysis:
(1) the classification of PDS, (2) the cost estimation for
health, investment risks and financial impact, and finally
(3) the estimation of the cost for the reliability improvement
of each system. To handle such situations, we have tried to
238 J.-E. Yang et al. / Reliability Engineering and System Safety 65 (1999) 229–238
get the bands of values for each variable instead of a single
value. Such bands are derived from two extreme values like
the upper and lower limits of each variable’s range.
Although, it does not show the band in which the real
optimum value exists, we can get some insights on the
ranges of system reliabilities.
In addition, as mentioned earlier, the reliabilities of
containment systems such as spray, hydrogen igniter, etc.
can be optimized if Level 2 PSA model is used to estimate
the health risk. In this case, containment event trees are to be
developed and the effect of the source term is to be evalu-
ated.
Even though, there are a number of issues to be solved,
the analyzed results show that the genetic algorithm can be
used as a useful decision-supporting tool to optimize the
design of a NPP. This approach can be also used not only
for the optimization of system reliability for a NPP but also
for the optimization of component reliability with the pre-
fixed reliability target of that system.
There is one thing that we have to keep in mind, that is,
the results obtained from the reliability allocation are not the
absolute values owing to the uncertainties from various
sources. Hence, it is meaningless to try to get absolute
numbers like the reliability target from reliability allocation.
Instead, we have to generate various alternatives under the
various assumptions, and we have to try to gain some insight
from those results. The genetic algorithm is a good tool that
provides us great flexibility to analyze sample cases under
various assumptions.
Acknowledgements
This research was supported by ‘‘The Mid- and Long
Term Nuclear R&D Program’’ of MOST (Ministry of
Science and Technology), Korea. The authors would also
like to thank the reviewers for their critical review and
constructive suggestions to improve the quality and read-
ability of this article.
References
[1] Yang XP, Kastenberg WE, Okrent D. Optimal safety goal allocation
for nuclear power plants. Reliability Engineering and System Safety
1989;25.
[2] Cho NZ, Papazoglou IA, Bari RA. Multi-objective programming
approach to reliability allocation for nuclear power plants. Nuclear
Science and Engineering 1987;95.
[3] Park CK, Bari RA. A consistent and cost-effective quantification of
containment performance criteria. Nuclear Technology 1988;81.
[4] Cho NZ, Papazoglou IA, Bari RA. A methodology for allocating
reliability and risk. NUREG/CR-4048, 1986.
[5] Jung WS, Cho NZ. Determination of performance criteria of safety
systems in a nuclear power plant via simulated annealing optimization
method. Ph.D. Thesis, KAIST, August 1993.
[6] Gokcek O,Temme MI, Derby SL. Risk allocation approach to reactor
safety design and evaluation. Proc. Topl. Mtg. Probabilistic Analysis
of Nuclear Reactor Safety, 1978.
[7] Knoll, A. Component cost and reliability importance for complex
system optimization. Proc. Int. ANS/ENS Topl. Mtg. Probabilistic
Risk Assessment, 1982.
[8] Koza JR. Genetic programming, Cambridge, MA: MIT Press, 1992.
[9] Painton L, Campbell J. Genetic algorithms in optimization of system
reliability. IEEE Transaction on Reliability 1995;44(2).
[10] Kumar A, Pathak RM, Gupta YP. Genetic algorithm based reliability
optimization for computer network expansion. IEEE Transaction on
Reliability 1995;44(1).
[11] Munoz A, Martorell S, Serradell V. Genetic algorithms in optimizing
surveillance and maintenance of components. Reliability Engineering
and System Safety 1997;57.
[12] EPRI NP-2529, Value-impact methodology for decision makers,
August, 1982.
[13] EPRI TR-102240, Evaluation of the safety benefits and costs of
proposed revisions to inservice testing requirements for pumps and
valves, September 1993.
[14] EPRI NP-3434, Value-impact analysis of selected safety modifica-
tions to nuclear power plants, March 1984.