You are on page 1of 2

DISCUSSIONS AND CLOSURES

Table 1. Values of R E ⫽H D 兩 D 共After Redefinition兲


Discussion of ‘‘Redundancy Index m⫽2 4 6 8 10 12 14 16
of Lifeline Systems’’ by Masaru Hoshiya
Ductile system 0.811 1.32 1.59 1.43 1.86 1.85 2.25 2.23
and Kinya Yamamoto 共cf. Table 1兲
September 2002, Vol. 128, No. 9, pp. 961–968. Brittle system 0.384 0.804 1.12 1.26 1.34 1.49 1.47 1.71
DOI: 10.1061/共ASCE兲0733-9399共2002兲128:9共961兲 共cf. Table 2兲

S. K. Au1
Downloaded from ascelibrary.org by New York University on 05/15/15. Copyright ASCE. For personal use only; all rights reserved.

1
Example 2
School of Civil and Environmental Engineering, Nanyang Technological
Univ., Singapore 639798. E-mail: cskau@ntu.edu.sg Consider the parallel link system as shown in Fig. 1 of the paper.
Assume m⭓2 and P D i 兩 D ⫽1/m for all i⫽1,...,m, then
The concept of redundancy has long existed in engineering design m⫺1
1 1 1 1
philosophy, but there has not been a common consensus on its
quantitative definition 共assuming it is quantifiable兲, especially in a
H D 兩 D ⫽⫺ 兺
i⫽1 m
log2 ⫺ log2 ⫽log2 m
m m m
(2)
probabilistic context where effects of system uncertainties are in-
corporated. In order for a redundancy index to be useful, its defi- and so
nition and resulting implications should be consistent with the H D兩D log2 m
R E⫽ ⫽ ⫽1 (3)
intuitive notion of redundancy. log2 m log2 m
The subject paper presents an interesting discussion on quan- This means that R E is always equal to 1, regardless of the number
tifying redundancy of engineering systems using conditional en- of links in the system. This conflicts with the intuition that in this
tropy and its applications. Some discrepancies in conclusion case a system with more links should have a higher redundancy.
about system redundancy, however, have been found when com- In fact, the conditional probability of system failure given at least
paring the proposed redundancy index R E with other existing in- one component has been damaged, which coincides with the re-
dexes through the numerical examples presented in the paper. A dundancy index proposed by De et al. 共1989兲, is given by
closer examination reveals that the discrepancies essentially stem
1
from the normalizing factor log2 m in the definition of R E . A P 共 D F兩 D 兲 ⫽ (4)
short conclusion of this discussion is that it may be more appro- m
priate to redefine R E without the normalizing factor, i.e., as R E which decreases as m increases, and hence, suggests an increase
⫽H D 兩 D . The problems associated with the normalizing factor in redundancy with m.
will be discussed through two examples. Note that for a given m the maximum value of H D 兩 D is log2 m,
and so the normalizing factor log2 m has the effect of comparing
the uncertainty of D with its maximum attainable value of a sys-
Example 1 tem with m links. This, however, causes problems in the use of R E
for comparing the redundancies of systems with a different num-
Consider the case when all the damage sets 兵 D 1 ,D 2 , . . . ,D m⫺1 其 ber of links, which is of concern when studying the benefits on
are absent so that D⫽D F . An example of this is a system with redundancy of introducing new members, for example. As can be
only one link. Thus, the system will not be operational whenever seen in Table 1 in the paper, as m increases, the value of R E tends
any one of its components fails. It is then clear that the system has to decrease, while on the contrary the value of
no redundancy, and a properly defined redundancy index should R⫽ P共operational兩at least one component fails兲 increases. In fact,
assume its minimum possible value. if R E were defined without the normalizing factor, which is then
In this case, in the context of Eq. 共8兲 in the paper, m⫽1, equal to the conditional entropy of D, i.e., R E ⫽H D 兩 D , then the
P D i 兩 D ⫽0 for i⫽1, . . . ,m⫺1, and P D F 兩 D ⫽1. By noting that the values of R E will show an increasing trend as the number of links
m increases 共see Table 1 in this discussion兲, and will be consistent
first sum in the numerator of Eq. 共8兲 vanishes 共using the usual
with the conclusions indicated by R Z and R.
interpretation in probability theory that 0⫻log2 0⫽0), one obtains
Finally, the discusser would like to point out that the insensi-
H D兩D ⫺1 log2 1 tivity of R Z is due to the relatively high reliability of the weakest
R E⫽ ⫽ (1) redundant component in the system that leads to a small uncer-
log2 m log2 1
tainty in the operational modes and renders the conditional en-
which is of the indeterminate form 0/0 since log2 1⫽0. If one tropy 共given the system is operational兲 relatively insensitive to the
interprets by canceling out the term log2 1 in Eq. 共1兲, then R E conditional uncertainty of the component failure modes 共given
⫽⫺1, which is absurd. One interpretation that leads to a reason- component damage has occurred but system is still operational兲.
able conclusion is to take the log2 1 in the numerator to zero first
before dividing it by the denominator, in which case R E ⫽0 and is Reference
consistent with intuition.
Although this example appears to be of an interpretation prob- De, R. S., Karamchandani, A., and Cornell, C. A. 共1989兲. ‘‘Study of
lem, the next example indicates that the normalizing factor indeed redundancy in near-ideal parallel structural systems.’’ Proc., 5th Int.
leads to problems in its meaning and use. Conf. on Structural Safety and Reliability, San Francisco, 975–982.

1104 / JOURNAL OF ENGINEERING MECHANICS © ASCE / SEPTEMBER 2003

J. Eng. Mech. 2003.129:1105-1105.


When plural existing systems need to be compared, the redun-
Closure to ‘‘Redundancy Index dancy differences among them can only be judged relatively. In
of Lifeline Systems’’ by Masaru Hoshiya such a case, it is not possible to make a final decision about
and Kinya Yamamoto whether or not all systems considered need disaster mitigation
September 2002, Vol. 128, No. 9, pp. 961–968. measures, based on values of a non-normalized redundancy
DOI: 10.1061/共ASCE兲0733-9399共2002兲128:9共961兲 index. As a conclusion, the authors have introduced a normalized
redundancy index R E .
Masaru Hoshiya, M.ASCE,1 and Kinya Yamamoto2 The discusser shows also a discrepancy by using an example
1
Professor, Dept. of Civil Engineering, Musashi Institute of Technology, in which R E is constant, taking a unity value regardless of the
1-28-1, Tamazutsumi Setagaya-ku, Tokyo 158-8557, Japan. E-mail: number of members m in a parallel system if P Di/D ⫽1/m for all i
mhoshiya@eng.musashi-tech.ac.jp from 1 to m. In this case, the redundancy index R E does not
2
Assistant Professor, Dept. of Civil Engineering, Kogyokusha College of increase with m, a behavior that seems to be illogical intuitively.
Technology, 5-14-2, Nishi-Gotanda Shinagawa-ku, Tokyo 141-0031, The values of P(D i ) presented in Tables 1 and 2 and the corre-
Japan. E-mail: yamamoto@kogyokusha.ac.jp
Downloaded from ascelibrary.org by New York University on 05/15/15. Copyright ASCE. For personal use only; all rights reserved.

sponding graphs presented in Fig. 4 make it clear that both par-


allel cable systems of ductile and brittle materials indicate chain
The authors would like to thank Professor S. K. Au very much for
failure phenomena when the number of components increases,
his constructive discussions on our paper.
though the severity of the phenomenon is different. It is indicated
The discusser suggests that the index R E proposed by the au-
that the degree of redundancy is more or less the same, at least for
thors will become a better index if only the numerator is used to
define R E . The reason, however, that we decided to normalize the systems with values of m greater than 6. This fact is understand-
conditional entropy into the range from 0 to 1 is our desire to ably attributed to the chain failure. On the other hand, the fact that
examine the degree of redundancy that a current system retains the ductile systems with m⫽2 – 6 presented in Fig. 4 show a trend
and to know the improvement on the redundancy when the sys- of decreasing R E should not be used for comparison purposes,
tem is strengthened. From the practical point of view, it is also since R E simply indicates the current redundancy relative to its
necessary that the redundancy index retains its applicability to possible maximum value within the system. It can be therefore
decision making in engineering system design and practice. deduced that the system reduces its relative redundancy due to
The discusser points out that the redundancy index R E does chain failure when the value of m value changes from 2 to 6.
not take the value of 0, but the value of 0 divided by 0, for the However, we would like to mention that the suggestion of the
case of a system with m⫽1. According to the discusser this indi- discusser that R E will become a more useful index only if the
cates a shortcoming in the definition of R E , since intuitively such normalizing factor log2(m) is discarded is indeed worthy of con-
a system should have no redundancy. This discrepancy is due to sideration. Especially because of the case when m⫽1 and R E
the normalization of the conditional entropy by the maximum becomes subsequently undefined, the authors are ready to inves-
value of log2(m). The authors believe, however, that the index tigate the topic again, regardless of the aforementioned answers.
becomes more useful when normalized, as it provides a better The authors further agree with the discusser’s reasoning regarding
way to understand the degree of redundancy. For example, con- the insensitivity of R Z , and believe that R E is superior to the
sider a system of m members having R E ⫽0.05. It can be easily other two indices R Z and R R , since the latter index R R unfortu-
determined that the redundancy is very low in this case, whereas nately cannot express the effect of chain failure occurrence as
it is very high if R E ⫽0.9, since R E can only take values between clearly shown in Tables 1 and 2 and Fig. 4.
zero and unity. If the index is not normalized and R E ⫽4.89, then Finally, the authors agree with the necessity to define the re-
what does this value mean regarding the degree of redundancy? dundancy index in a mathematically valid and consistent way,
An absolute value without units tells us nothing in this case. keeping at the same time the intuitive notion of redundancy.

JOURNAL OF ENGINEERING MECHANICS © ASCE / SEPTEMBER 2003 / 1105

J. Eng. Mech. 2003.129:1105-1105.

You might also like