Uncertainty Management

Alfs T. Berztiss1;2
1
University of Pittsburgh, Department of Computer Science, Pittsburgh PA 1526
USA
2
SYSLAB, University of Stockholm, Sweden
e-mail: [email protected]

Abstract. The purpose of this fairly nontechnical introduction to un-

certainty management is to identify various forms of uncertainty, and
to survey methods for managing some of these uncertainties. Our em-
phasis is on topics that may not be familiar to software engineers or, to
a lesser extent, to knowledge engineers. These topics include Bayesian
estimation, fuzziness, time Petri nets, rough sets, belief and evidence,
and possibility theory. Uncertainty management has been studied in the
contexts of information systems and of articial intelligence. We attempt
to present a balanced view of the contributions from both areas.
Keywords : belief, fuzziness, possibility, probability, rough set, time Petri
net, uncertainty.

1 Introduction
Most decisions in life have to be made under uncertainty. This applies also to
decisions made by software systems. But uncertainty can be managed, that is,
it can be allowed for in a systematic and disciplined way. The purpose of this
chapter is to survey approaches to the management of uncertainty. Section 2 is
a catalog of dierent types of uncertainty. Section 3 is a brief survey of basic
discrete probability and stochastic inference, including Bayesian inference. Fuzzy
sets are introduced in Sec. 4, and fuzzy logic in Sec. 5. In Sec. 6 we outline the
operation of a fuzzy control system, including defuzzication.
Time Petri nets allow the introduction of uncertainty into the time aspects
of real-time processes. An application of time Petri nets is discussed in Sec. 7.
Similarity gives rise to a particular type of uncertainty | given a set of objects
that are to be grouped into dierent classes, what criteria should be used to put
a particular object into one of the classes? The theory of rough sets, which is
looked at in Sec. 8, is one way of dealing with this problem. Section 9 is concerned
with belief, and Sec. 10 with possibility theory. In the interests of readability we
have avoided sprinkling the text with references, putting nearly all bibliographic
pointers into Sec. 11.

2 Types of uncertainty
In [1], which is probably the best reference on various aspects of uncertainty, un-
certainty issues are partitioned into those that deal with fuzziness and those that
deal with ambiguity. Fuzziness is taken to deal with lack of distinctions. Ambi-
guity problems are further partitioned into discord problems and nonspecicity.
Discord is a disagreement when a choice between dierent alternatives is possi-
ble; nonspecicity relates to situations in which some such possible alternatives
are not taken into account. In [2] we set up a ner classication scheme, which
we present here in a somewhat expanded form. For quite a few of the uncertainty
types there is no denite procedure for dealing with them, but awareness of their
existence will at least cause us to make allowances for them.
Inconsistency. This arises when con icting solutions to a problem are being
advanced, notably in dietary recommendations and in economics. For example,
both lowering and raising taxes may be suggested as a cure for reducing budget
decits. Software requirements can state that costs are to be low, and reliability
is to be high | satisfaction of one of these objectives normally means that the
other cannot be satised.
Permanent exceptions. Emus and penguins are exceptions to \All birds y".
This problem is easily solved by a more rened classication of the entities under
consideration.
Temporary exceptions. Suppose that all vice-presidents are to have oces
on a particular oor of the head oce buiding. If Ms. Smith has been made a
vice-president, but has not yet made a move to this oor, we have a temporary
exception.
Limited validity. Although Jack may not own a car, there are times when he
\has" a car, borrowed from his parents. It is not clear to what extent one \owns"
a heavily mortgaged house.
Multiple options. People often have several addresses, used for dierent pur-
poses. There may be dierent options on how to y from place A to place B |
cost or convenience may decide which of them is selected.
Nondeterminism. If a software system interacts with its environment the
behavior of the system is not fully predictable from knowledge of its initial
state.
Obscurity. A general trend in data may be obscured by temporal variations,
e.g., monthly income of a ski resort.
Fuzziness. We designate Mr. Smith as tall, but tallness is not a well-dened
property. We interpret a property as fuzzy if a precise measurement of this
property can be obtained in principle. Examples of fuzzy terms: cold, old, loud.
Vagueness. In [1] vagueness is considered a sub-category of fuzziness. We
consider vagueness to be distinct from fuzziness. In contrast to fuzzy terms, we
call those terms vague for which no measurement process can exist. In \People
feel uncomfortable when it is hot" the term \hot" is fuzzy, but \uncomfortable"
is vague | we have no dependable way of measuring discomfort. Part of the
research on uncertainty should be aimed at reducing vagueness by developing
new measurement processes.
Faults. An electrical utility meter breaks down. In some circumstances it
would help to know when this happened.
 Either/or uncertainty. If this meter registers the same value over a period of
time, it may be broken, but not necessarily. The rumor \Company X will show
a loss this quarter" may be true or false, but we do not know which.
Cause/eect uncertainty. An eect is observed that could be due to dierent
causes, singly or jointly. This is quite common in medical diagnosis. Analysis
methods based on fault trees and Bayesian networks have been developed to
deal with such situations.
No knowledge as negation. In logic programming, if the fact X is not in its
knowledge base, not(X) is assumed true. For example, if the knowledge base does
not contain \Mr. Khachaturian is married," and this cannot be derived from,
say, \The husband of Mrs. Khachaturian is Mr. Khachaturian," the answer to
the query \Is Mr. Khachaturian married?" will be \no".
Null values. This problem has been extensively studied by the data base
community. It has two aspects. One is representational: when a value is missing,
how should this be indicated? A missing value may exist, but be unknown (John's
age), not exist at all (temperature data for a given place and date because no
reading was taken), or be inapplicable (the name of the spouse of an unmarried
person). Dierent null-markers may be used to dierentiate between these three
types of null values, but at times there may be uncertainty as to which marker is
appropriate. The other aspect relates to query answering. If the data base does
not contain data needed to respond to a query, an approximate answer may be
given. Here the uncertainty relates to similarity: how are we to determine what
approximate answer is appropriate?
Interpretation uncertainty. At one time the author worked in an insurance
oce where three dierent interpretations of \age" were in eect, namely age
last birthday, age nearest birthday, and age next birthday. In the United States a
weather forecast may state that tomorrow there is a 60% chance of precipitation.
What exactly does this mean?
Rounding. This is interpretation uncertainty as it applies to numerical data.
Is USD 5,000,000 an exact amount or is it a rounded estimate? As the purchase
price for a mansion it is probably exact; as damage estimate due to ooding it
is likely to be a rounded gure.
Noisy data. An experimentally determined value is nearly always uncertain
to some degree, e.g., the speed of light. Also, data may become polluted during
transmission, requiring the use of error-detecting codes.
Computational round-o. As an illustration, if one keeps adding up oating-
point representations of the number 1, at some point the sum no longer changes.
Dierent results are obtained depending on whether a series is added from the
small values rst or the large values rst.
Computational result uncertainty. This can take two forms. First, the pro-
gram is correct, but the result it gives need not be the required solution. The
aim of nonlinear optimization is to nd the maximum of an objective function.
A two-variable function f(x; y) can be pictured as a three-dimensional hilly ter-
rain. Hill climbing methods will nd the top of a hill, i.e., a solution, but there is
no guarantee that this is the highest hill. The problem becomes more and more
dicult as the number of variables goes up. Under the second form we have no
way of telling whether a program is correct. The whole point of many programs
is to arrive at results that cannot be obtained in any other way. Hence there is
really no way of testing such a program, and a result will be accepted as long as
it \looks reasonable".
Partial knowledge. Six prisoners escape. Their names are known. We also
know that they have split into two groups. However, we may not know who is
in each group, or even the size of each group.
Trends. An upward trend has been observed in global temperatures. Is this
due to chance or is there an underlying reason for it?
Context-dependence. Many terms cannot be fully understood unless the con-
text is known in which they arise. For example, a warm day near the Arctic
Circle is likely to be cooler than a cool day near the Equator. Since the concepts
\warm" and \cold" are fuzzy, the example shows that care must be taken in the
interpretation of fuzzy terms. Although we take the boiling point of water to be
100C, at high elevations it is actually lower.

3 Probabilistic concepts
Probability theory was the earliest attempt to deal with uncertainty in a dis-
ciplined quantitative manner. Here we can do no more than give a very brief
review of basic discrete probability. In particular, we introduce Bayesian esti-
mation, which is being applied extensively in articial intelligence to deal with
uncertainty.
Consider an experiment that can have k possible outcomes or events a1 ,
a2, ..., ak . This experiment is performed n times, and the counts of observed
outcomes are n1 , n2, ..., nk . Then p(ai ) = ni =n is the observed probability of
outcome ai. We have
Xk Xk X k
p(ai ) = ni=n = n1 ni = 1:
i=1 i=1 i=1
The probability of the disjunction of outcomes a1, a2, ..., at is the sum of
their respective probabilities; the probability of the conjunction of the outcomes
is the product of their probabilities:
X t t
Y
p(ai ); p(ai ):
i=1 i=1
Thus the probability of the throw of a die resulting in a 3 or a 6 is 1/6 + 1/6
= 1/3, and the probability of the rst throw of a die resulting in a 3 and the
second throw in a 6 is 1/6  1/6 = 1/36.
Conditional probability determines the probability of an event, say A, given
that another event, say B, has occurred:
P (AjB) = P(A \ B) :
P(B)