You are on page 1of 5

Advance Topics in Expert Systems

Uncertainty

In real life situation, often the information is not ‘black and white’ or ‘yes and
no’ type. We cannot always be 100-percent sure about the value of an
attribute. In such situations involving ‘uncertainty’, we use confidence factors
in developing the expert systems. In expert systems, the ‘uncertainty’ can be
one of the two types:

(a) Uncertainty in Conclusions

Also called ‘output uncertainty’, is the situation when we are not able to
conclude with 100-percent confidence. In other words, we cannot assign the
value to an attribute in a conclusion with complete confidence. Consider the
following example:

Rule 1: IF temperature = moderate


AND humidity = high
THEN rain = yes

For the combination of the premise attribute values in the above rule, it may
not be possible to assign the value of ‘yes’ to the conclusion attribute ‘rain’.

Let us say that we have historical records of two years. Based on these
records, let us say that we have the premise situation for 100 days but it
rained on only 80 days (out of 100 days). Therefore, we can now have the
following rule:

Rule 1a: IF temperature = moderate


AND humidity = high
THEN rain = yes (cf 0.80)

Clearly, such a rule assigns a certain ‘confidence factor’ to the conclusion


clause.

(b) Uncertainty in Premises

Also called ‘input uncertainty’, is a situation when the user is not 100-percent
about the value of a particular attribute appearing in the premise of a rule.

1
Consider the following query:

Does the patient have severe stomach cramps?

Obviously, the expected answer is either ‘yes’ or ‘no’.

o Now, if the pain is so sever that the patient cannot even talk or sit
properly then the answer is ‘yes’.
o Or, if the patient feels completely comfortable and no stomach cramps,
then the answer is ‘no’.
o But there can be a situation when the patient may have pain but it is not
so severe.
o The patient may be able to identify the intensity on a scale of 0 to 10.

Then, an appropriate question might be as follows:

Indicate the intensity of any stomach cramps (0 to 10) …….

Handling Uncertainty

Handling of the uncertainty in expert systems is a controversial issue and


there is no consensus. This is because the confidence factors are not truly the
probabilities as they may sometimes be developed based on judgment,
experience, and intuition that are not subjected to the laws of probability.
Consider the following situation:

Rule 1: IF <Premise 1: cf 0.80>


AND <Premise 2: cf 0.40>

This is the premise of a rule consisting of two clauses having certain level of
confidence. What should be the confidence factor of the combined (or whole
premise)

o Should it be the minimum of the two confidence factors? (0.40)


o Should it be the maximum of the two confidence factors? (0.80)
o Should it be the product of the two confidence factors? (0.32)
o Should it be the average of the two confidence factors? (0.60)

2
Often, which approach should be used, depends upon the kind of application.
Therefore, many expert system software provide various options to calculate
combined confidence factor.

Another problem in dealing with uncertainty in expert systems using laws of


probability is estimating many types of probabilities that may be either too
difficult, impractical, or even impossible. Consider the Bayes’ Formula

Bayes’ Formula:

Bayes’ formula can be used to determine the probability of a conclusion (C),


given certain evidence or facts (f) as follows:

P (C | f) = {P (f | C) * p(C)}/ p(f)
p(f) = P (f | C) * p(C) + P (f | ~C) * p(~C)
Where

P (C | f) = probability of conclusion C given facts f


P (f | C) = probability of facts f given conclusion C
p(f) = unconditional probability of facts f
p(C) = unconditional probability of conclusion C
p(~C) = unconditional probability of NOT conclusion C
P (f | ~C) = probability of facts f given NOT conclusion C

Consider the problem of classifying the computer disks as good or defective.

Rule 1: IF unusual noise = yes


THEN disk drive status = defective

For this example,

P (C | f) is the probability of a defective disk drive given unusual noises


P (f | C) is the probability of unusual noises given a defective disk drive
p(f) is the probability that any disk drive is noisy
p(C) is the probability that any disk drive is defective
p(~C) is the probability that any disk drive is not defective
P (f | ~C) is the probability of unusual noises from a good disk drive

It should be clear why it is so difficult to employ formal probability theory for


the development of confidence factors in production rules.

3
Uncertainty through Fuzzy Sets

Uncertainty in expert systems is sometimes dealt with using the fuzzy set
theory. Fuzzy set theory was proposed by Zadeh, L. (1965) to deal with
problems where a simple yes or no is inadequate.

Example: Consider the query – Is the person tall?

o If the person is adult male of height 7 feet, the response is ‘definitely


yes’
o If the person is adult male of height 3’ 6”, the response is ‘definitely no’
o If the person is 5 feet 11 inches, the response is ???

Fuzzy Membership Function: A fuzzy membership function is defined as a


function with value in the range of 0 and 1 to describe the confidence in
answering a query or characterizing a particular object.

o In the above example, the fuzzy membership function is nothing but the
partial degree of tallness of a person.
o For example, a person of height 5 feet 11 inches may have a fuzzy
membership function value of 0.75 OR his degree of tallness is 0.75.

Certain Rules to Evaluate Conclusions using Fuzzy Logic

o If premises are connected with the logical operator AND, use the
minimum of the fuzzy membership function values associated with the
premises to determine the composite value of the premise.
o If premises are connected with the logical operator OR, use the
maximum of the fuzzy membership function values associated with the
premises to determine the composite value of the premise.
o If the fuzzy membership function value of a premise i is given as fv(i)
then NOT fv(i) = 1-fv(i).

4
Confidence Factor Union Method

When we have two different rules concluding about an attribute with different
confidence factors cf1 and cf2, we can use the union method to find combined
confidence factor. Then the combined confidence in the final conclusion is
given by the following rule:

C(cf) = cf1 + cf2 – cf1*cf2

The union method can be chained when we have more than two rules.

Example: cf1 = 0.2; cf2 = 0.5; cf3 = 0.5

C(cf1&2) = 0.2+0.5 – 0.2*0.5 = 0.6

C(cf) = 0.5 + 0.6 – 0.5*0.6 = 0.80

Clearly, as the number of rules assigning a value to a conclusion attribute


increases, the overall confidence in the conclusion also increases. This is more
intuitive as compared to using either a minimum or maximum approach.

Advanced Topics in Expert Systems

Interfaces or Bridges

Interfaces or bridges are the capabilities of expert system software that allows
the control of inference engine to go out of the expert system temporarily
during reasoning to perform a specific task that can help in making a final
decision. Most of the expert system software come with in-built interface
capability today. Some of the interfaces or bridges available include the
following capabilities

o Graphics Interface for visual interpretation


o Database Interface for input and output
o Simulation Interface to simulate physical processes
o Statistical Interface to calculate statistics
o Mathematical Interfaces for optimization, LP, DP, solution of OE, PDEs
etc.

You might also like