Abstract
Many quantities of interest in ethics analysis: reasonable person, worthwhile goal, serious
consequences, responsible risk, likelihood of failure, reliability of predictions, special
responsibilities (personal, membership, skill-based) – are well-represented by linguistic
variables denoted by fuzzy membership functions. This is a natural extension of fuzzy logic
applications in high-consequence risk analysis. The common property is information that is not
as well portrayed by other methods of representing uncertainty.
I. Introduction
I.a. Decision Support
In many important but complex decision-making processes, the use of decision support systems
is applied to augment human reasoning capabilities. Reasons for this include the complexity of
the process, accessibility to data, record-keeping, and the ability to leverage related
computational capabilities, such as visualization, simulation, and the Internet. Some of the areas
where such tools are routinely applied are of great interest in ethics – for example, high-
consequence risk assessment.
Examples of risk assessment can be illustrative, and will be therefore explored a bit further.
Transportation of nuclear materials must be done with great care, due to the possibility of
catastrophic accidental or deliberate intervention. In this case, the decision support tools serve to
sort out the complexity due to proliferation of scenarios and uncertainty associated with each
critical link in the process. A simpler example occurs in the insurance industry, where the
pooling of risks allows use of the Law of Large Numbers to obtain reliable estimates of risk-
reward tradeoffs. This example is quickly made difficult again by considering the case of
reinsurance. The reinsurance industry exists to shield insurers from unsafe aggregation of risks,
for example, too many homeowners’ insurance policies in one large location, potentially subject
to hurricane damage or similar disasters, such as Houston. In such cases, of which the preceding
example is a relatively simple one, assessing the risk-reward tradeoff is fraught with difficulty,
It would appear that such tools would be highly useful in analysis of ethics questions. Many
ethical issues involve sufficient complexity, that analysis tools would be called for in most other
disciplines involving similar complexity. Perhaps the primary reason for the lack of application
of such tools to date is the fuzzy nature of the variables of consideration. We will return to this
point later, but for the moment, will introduce some important aspects of Engineering Ethics to
focus the discussion.
It is illustrative to consider the IEEE Code of Ethics. This emphasizes: responsibility for safety,
conflicts of interest, honesty of claims and estimates, rejection of bribery, improving
understanding, maintaining competence, credit and criticism, fair treatment and avoiding injury,
and assisting colleagues. The issues most affected by this paper are safety, and claims and
estimates.
It is also illustrative to look at several concepts from Engineering Ethics that are affected by
imprecision or uncertainty. The items below are just a sampling of such concepts pulled from
[2].
Risk
“The probability that a given course of action will produce some harm, multiplied by degree of
harm, defines risk … (All quotes cited retain emphasis in original.) Probability will be
discussed at more length below, but note that degree of harm is already a fuzzy concept.
Conflict of Interest
Conflict of interest is frequently defined to apply to the conflict between an individual’s private
interests as opposed to their professional obligations. It is the definition most frequently applied
in engineering contexts. However, this is not the only possible definition. Lawyers, for
example, may have a conflict of interest even if their personal interests are kept separate, due to
conflicting professional obligations. This is due to the adversarial nature of litigation. Whether
Even a bare-bones introduction to Fuzzy Logic is beyond the scope of this paper, but the most
salient points, assumed in the discussion below, are summarized. Fuzzy logic is based on the
concept of degree of membership in a set, as quantified by a membership function, m. Standard
set theory is reduced to a special case by considering only values of m = 0 or 1. Any levels in
between are partial degrees of membership. It is possible to use a threshold function, or a–cut,
to map fuzzy sets back onto non-fuzzy or crisp sets. It is also possible to represent numerical
quantities by fuzzy membership functions. These fuzzy numbers would typically have certain
properties, such as unimodality and symmetry.
It is illustrative to compare fuzzy logic with the most popular alternative representation of
uncertainty, probability. Two examples will clarify this comparison, both due to Bezdek. [3,4]
In one case, imagine a traveler, nearly dying of thirst, who finds two bottles of liquid. One is
marked: “Potable with probability 0.9.” The other is marked: “Potable with membership
function 0.9.” It is argued that the second bottle is preferable, in that the first incurs a 10%
chance of sickness or death, whereas the second is certain to be drinkable. As another example,
the author is (at this writing) neither bald, nor not bald. Probability is not involved at all,
because the uncertainty is not due to randomness or to unknown data. The uncertainty is due to
vagueness in language, something that humans routinely deal with, but which is more difficult
for formal reasoning systems and computers.
It is also important to point out that standard, or crisp, set theory can be viewed as a special case
of fuzzy set theory, and that probability can be viewed as a special case of fuzzy logic.
Perhaps the most critical assumptions are statistical independence and point estimates of
probabilities. Statistical independence assumptions can be relaxed, at a cost of increasing
complexity. Even when this is not done, conclusions will often be modified after-the-fact due to
known limitations of the independence assumptions. This is one way of dealing with the
insurance risk problem. In high consequence applications, the probability estimates themselves
might be the most critical problem. For example, what is the probability of a critical
In [6], Cooper argues that such probabilities are best represented by fuzzy numbers. This allows
a range of reasonable values to take the place of a single wild guess. He demonstrated how to
accomplish fuzzy fault-tree analysis by replacing probabilities by fuzzy numbers in this manner.
In [7], some reformulation of fuzzy logic approaches is undertaken to facilitate and simplify a
similar technique.
In the lightning example, the statement, “The probability of the system being struck by lightning
is 1 x 10-8,” would be replaced by “The probability of the system being struck by lightning is
represented by the following membership function.” The function specified might be trapezoidal
in shape, with m = 0 for probability values below 1 x 10-15 or for probability values above 1 x 10-4
; and m = 1 for values between 1 x 10-12 and 1 x 10-6 . The value of m might vary linearly or
log-linearly for the regions where it transitions between zero and one.
A fault tree built on the above technique could have a combination of fuzzy and crisp probability
specifications. It turns out that their combinations are well-defined. The output of such a system
is a fuzzy membership function for the risk assessment, with a range similar to that described
above. Knowing a range of reasonable assessments of risk may, in many cases, be superior to
knowing a point estimate. Recall the design analogy of ethical analysis. There might not be a
unique answer to an ethical query, and the range of reasonable interpretations of risk could be
part of the reason. Note that it is not sufficient to simply choose the most conservative risk
estimate in all cases. Most risk mitigation incurs costs, including costs that could be better spent
to mitigate alternative risks. In [8], Unger points out that technology decisions are made based
on an analysis of tradeoffs between benefits and risks. This is true, as this paper has argued, of
non-technical decisions as well. This analysis is obviously affected by where in the range of
reasonable estimates a risk is assessed. Similar observations could be applied to benefits. Tools
that consider the whole range of reasonable assessments, and return a whole range of reasonable
results, can therefore be of great value in this analysis.
Of course, a range of probability values can be analyzed completely in the context of probability.
However, these techniques are most appropriate when enough data is present for the law of large
numbers to make the estimates meaningful. In high-consequence applications, this is
(fortunately!) typically not the case. However, this does not mean that bounds can’t be set for
the reasonableness of probability estimates. However, these bounds will typically be determined
by human expert opinion, and thereby be better modeled by fuzzy membership functions.
IV. Conclusion
It is argued that fuzzy logic is an appropriate tool for engineering ethics analysis, particularly
situations involving PRA, or computing with words. The advantages and disadvantages of
probability vs. fuzzy logic were addressed, and the technique of using fuzzy membership
functions to achieve a representation of the range of reasonable probability estimates, in the
absence of sufficient supporting data for numerical approaches, was argued.
V. Acknowledgements
Partial support of this work by the National Science Foundation, Sandia National Laboratories,
and the Mary K. Finley Missouri Endowment are gratefully acknowledged.
VI. References
[1] For an illuminating brief discussion, see:
Berkshire Hathaway 2002 Annual Report, Warren Buffett, available online,
http://www.berkshirehathaway.com/letters/2002pdf.pdf , pp.7-10.
[2] Ethics in Engineering Practice and Research, Caroline Whitbeck, Cambridge University Press, 1998.
[3] “Editorial: fuzzy sets – what are they and why?,” Bezdek, Jim C., IEEE Trans. on Fuzzy Systems, Vol. 2, No. 1,
February 1994.
[4] “The thirsty traveler visits Gamont: a rejoinder to ‘Comments on fuzzy sets – what are they and why?,” Bezdek,
Jim C., IEEE Trans. on Fuzzy Systems, Vol. 2, No. 1, February 1994, pp. 43-45.
[5] “Taking Risks: Courage and Negligence,” Caroline Whitbeck, UMR Philosophy Seminar Series, April 9, 2003.
[6] “Fuzzy-algebra uncertainty analysis of abnormal-environment safety assessment,” J. Arlin Cooper, Journal of
Intelligent Fuzzy Systems, Vol. 2, No. 4, pp. 337-445, 1994.
[7] “A Theory of Independent Fuzzy Probability for System Reliability,” James Dunyak, Ihab W. Saad, and Donald
C. Wunsch, IEEE Trans. on Fuzzy Systems, Vol. 7, No. 2, April 1999, pp. 286-294.
[9] “Fuzzy logic = computing with words,” Lotfi A. Zadeh, IEEE Trans. on Fuzzy Systems, Vol. 4, No. 2, May
1996, pp. 103-111.
[10] Fuzzy Thinking: The New Science of Fuzzy Logic, Bart Kosko, Hyperion, 1993.
[11] Fuzzy Future: From Society and Science to Heaven on a Chip, Bart Kosko, Harmony Books, 1999.
DONALD C. WUNSCH II
Donald Wunsch II (Senior Member IEEE) received the Ph.D. in Electrical Engineering and the M.S. in Applied
Mathematics from the University of Washington in 1991 and 1987, the B.S. in Applied Mathematics from the
University of New Mexico in 1984, and completed a Humanities Honors Program at Seattle University in 1981.
Since July 1999, he is the Mary K. Finley Missouri Distinguished Professor of Computer Engineering in the
Department of Electrical and Computer Engineering, University of Missouri - Rolla. He heads the Applied
Computational Intelligence Laboratory and also has a joint appointment in Computer Science. Previously, he was
Associate Professor of Electrical and Computer Engineering, and Computer Science, at Texas Tech University.
Prior to joining Tech in 1993, he was Senior Principal Scientist at Boeing, where he invented the first optical
implementation of the ART1 neural network, featured in the 1991 Boeing Annual Report, and other optical neural
networks and applied research contributions. He has also worked for International Laser Systems and Rockwell
International, and consulted for Sandia Labs, White Sands Missile Range, Texas Tech, Boston University, Battelle
Pacific Northwest Labs, and Accurate Automation Corporation. Research activities include adaptive critic designs;
neural network pattern analysis, optimization, forecasting and control; computer security; bioinformatics; financial
engineering; fuzzy risk assessment for high-consequence surety; intelligent agents; graph theory; quantum logic; and
Go. He is heavily involved in research collaborations with former Soviet scientists. He is an Academician in the
International Academy of Technological Cybernetics, and in the International Informatization Academy; and is
recipient of the Halliburton Award for excellence in teaching and research at Texas Tech, and a National Science
Foundation CAREER Award. He is a Registered Professional Engineer in the State of Texas, and a member of the
International Neural Network Society, Association for Computing Machinery, American Society of Engineering
Educators, Phi Kappa Phi, a life member of the American Association of Artificial Intelligence, a life member of
Sigma Xi, and previously served as an Associate Editor of the IEEE Transactions on Neural Networks and voting
member of the IEEE Neural Network Council. He was elected to the International Neural Network Society’s Board
of Governors in 2001, and was Program Co-Chair of the IEEE/INNS International Joint Conference on Neural
Networks (IJCNN) ’02 and General Chair of IJCNN ’03. He has over 200 publications in the aforementioned fields
of interest.