[4] What is a fuzzy expert system?

Date: 21-APR-93

A fuzzy expert system is an expert system that uses a collection of
fuzzy membership functions and rules, instead of Boolean logic, to
reason about data. The rules in a fuzzy expert system are usually of a
form similar to the following:

    if x is low and y is high then z = medium

where x and y are input variables (names for know data values), z is an
output variable (a name for a data value to be computed), low is a
membership function (fuzzy subset) defined on x, high is a membership
function defined on y, and medium is a membership function defined on z.
The antecedent (the rule's premise) describes to what degree the rule
applies, while the conclusion (the rule's consequent) assigns a
membership function to each of one or more output variables.  Most tools
for working with fuzzy expert systems allow more than one conclusion per
rule. The set of rules in a fuzzy expert system is known as the rulebase
or knowledge base.

The general inference process proceeds in three (or four) steps. 

1. Under FUZZIFICATION, the membership functions defined on the
   input variables are applied to their actual values, to determine the
   degree of truth for each rule premise.

2. Under INFERENCE, the truth value for the premise of each rule is
   computed, and applied to the conclusion part of each rule.  This results
   in one fuzzy subset to be assigned to each output variable for each
   rule.  Usually only MIN or PRODUCT are used as inference rules. In MIN
   inferencing, the output membership function is clipped off at a height
   corresponding to the rule premise's computed degree of truth (fuzzy
   logic AND). In PRODUCT inferencing, the output membership function is
   scaled by the rule premise's computed degree of truth.

3. Under COMPOSITION, all of the fuzzy subsets assigned to each output
   variable are combined together to form a single fuzzy subset 
   for each output variable.  Again, usually MAX or SUM are used. In MAX
   composition, the combined output fuzzy subset is constructed by taking
   the pointwise maximum over all of the fuzzy subsets assigned tovariable
   by the inference rule (fuzzy logic OR).  In SUM composition, the
   combined output fuzzy subset is constructed by taking the pointwise sum
   over all of the fuzzy subsets assigned to the output variable by the
   inference rule.

4. Finally is the (optional) DEFUZZIFICATION, which is used when it is
   useful to convert the fuzzy output set to a crisp number.  There are
   more defuzzification methods than you can shake a stick at (at least
   30). Two of the more common techniques are the CENTROID and MAXIMUM
   methods.  In the CENTROID method, the crisp value of the output variable
   is computed by finding the variable value of the center of gravity of
   the membership function for the fuzzy value.  In the MAXIMUM method, one
   of the variable values at which the fuzzy subset has its maximum truth
   value is chosen as the crisp value for the output variable.

Extended Example:

Assume that the variables x, y, and z all take on values in the interval
[0,10], and that the following membership functions and rules are defined:

  low(t)  = 1 - ( t / 10 )
  high(t) = t / 10

  rule 1: if x is low and y is low then z is high
  rule 2: if x is low and y is high then z is low
  rule 3: if x is high and y is low then z is low
  rule 4: if x is high and y is high then z is high

Notice that instead of assigning a single value to the output variable z, each
rule assigns an entire fuzzy subset (low or high).

Notes:

1. In this example, low(t)+high(t)=1.0 for all t.  This is not required, but 
   it is fairly common.

2. The value of t at which low(t) is maximum is the same as the value of t at
   which high(t) is minimum, and vice-versa.  This is also not required, but
   fairly common.

3. The same membership functions are used for all variables.  This isn't
   required, and is also *not* common.


In the fuzzification subprocess, the membership functions defined on the
input variables are applied to their actual values, to determine the
degree of truth for each rule premise.  The degree of truth for a rule's
premise is sometimes referred to as its ALPHA.  If a rule's premise has a
nonzero degree of truth (if the rule applies at all...) then the rule is
said to FIRE. For example,

x       y       low(x)  high(x) low(y)  high(y) alpha1  alpha2  alpha3  alpha4
------------------------------------------------------------------------------
0.0     0.0     1.0     0.0     1.0     0.0     1.0     0.0     0.0     0.0
0.0     3.2     1.0     0.0     0.68    0.32    0.68    0.32    0.0     0.0
0.0     6.1     1.0     0.0     0.39    0.61    0.39    0.61    0.0     0.0
0.0     10.0    1.0     0.0     0.0     1.0     0.0     1.0     0.0     0.0
3.2     0.0     0.68    0.32    1.0     0.0     0.68    0.0     0.32    0.0
6.1     0.0     0.39    0.61    1.0     0.0     0.39    0.0     0.61    0.0
10.0    0.0     0.0     1.0     1.0     0.0     0.0     0.0     1.0     0.0
3.2     3.1     0.68    0.32    0.69    0.31    0.68    0.31    0.32    0.31
3.2     3.3     0.68    0.32    0.67    0.33    0.67    0.33    0.32    0.32
10.0    10.0    0.0     1.0     0.0     1.0     0.0     0.0     0.0     1.0


In the inference subprocess, the truth value for the premise of each rule is
computed, and applied to the conclusion part of each rule.  This results in
one fuzzy subset to be assigned to each output variable for each rule.

MIN and PRODUCT are two INFERENCE METHODS or INFERENCE RULES.  In MIN
inferencing, the output membership function is clipped off at a height
corresponding to the rule premise's computed degree of truth.  This
corresponds to the traditional interpretation of the fuzzy logic AND
operation.  In PRODUCT inferencing, the output membership function is
scaled by the rule premise's computed degree of truth.

For example, let's look at rule 1 for x = 0.0 and y = 3.2.  As shown in the
table above, the premise degree of truth works out to 0.68.  For this rule, 
MIN inferencing will assign z the fuzzy subset defined by the membership
function:

    rule1(z) = { z / 10, if z <= 6.8
                 0.68,   if z >= 6.8 }

For the same conditions, PRODUCT inferencing will assign z the fuzzy subset
defined by the membership function:

    rule1(z) = 0.68 * high(z)
             = 0.068 * z

Note: The terminology used here is slightly nonstandard.  In most texts,
the term "inference method" is used to mean the combination of the things
referred to separately here as "inference" and "composition."  Thus
you'll see such terms as "MAX-MIN inference" and "SUM-PRODUCT inference"
in the literature.  They are the combination of MAX composition and MIN
inference, or SUM composition and PRODUCT inference, respectively.
You'll also see the reverse terms "MIN-MAX" and "PRODUCT-SUM" -- these
mean the same things as the reverse order.  It seems clearer to describe
the two processes separately.


In the composition subprocess, all of the fuzzy subsets assigned to each
output variable are combined together to form a single fuzzy subset for each
output variable.

MAX composition and SUM composition are two COMPOSITION RULES.  In MAX
composition, the combined output fuzzy subset is constructed by taking
the pointwise maximum over all of the fuzzy subsets assigned to the
output variable by the inference rule.  In SUM composition, the combined
output fuzzy subset is constructed by taking the pointwise sum over all
of the fuzzy subsets assigned to the output variable by the inference
rule.  Note that this can result in truth values greater than one!  For
this reason, SUM composition is only used when it will be followed by a
defuzzification method, such as the CENTROID method, that doesn't have a
problem with this odd case. Otherwise SUM composition can be combined
with normalization and is therefore a general purpose method again.

For example, assume x = 0.0 and y = 3.2.  MIN inferencing would assign the
following four fuzzy subsets to z:

      rule1(z) = { z / 10,     if z <= 6.8
                   0.68,       if z >= 6.8 }

      rule2(z) = { 0.32,       if z <= 6.8
                   1 - z / 10, if z >= 6.8 }

      rule3(z) = 0.0

      rule4(z) = 0.0

MAX composition would result in the fuzzy subset:

      fuzzy(z) = { 0.32,       if z <= 3.2
                   z / 10,     if 3.2 <= z <= 6.8
                   0.68,       if z >= 6.8 }


PRODUCT inferencing would assign the following four fuzzy subsets to z:

      rule1(z) = 0.068 * z
      rule2(z) = 0.32 - 0.032 * z
      rule3(z) = 0.0
      rule4(z) = 0.0

SUM composition would result in the fuzzy subset:

      fuzzy(z) = 0.32 + 0.036 * z


Sometimes it is useful to just examine the fuzzy subsets that are the
result of the composition process, but more often, this FUZZY VALUE needs
to be converted to a single number -- a CRISP VALUE.  This is what the
defuzzification subprocess does.

There are more defuzzification methods than you can shake a stick at.  A
couple of years ago, Mizumoto did a short paper that compared about ten
defuzzification methods.  Two of the more common techniques are the
CENTROID and MAXIMUM methods.  In the CENTROID method, the crisp value of
the output variable is computed by finding the variable value of the
center of gravity of the membership function for the fuzzy value.  In the
MAXIMUM method, one of the variable values at which the fuzzy subset has
its maximum truth value is chosen as the crisp value for the output
variable.  There are several variations of the MAXIMUM method that differ
only in what they do when there is more than one variable value at which
this maximum truth value occurs.  One of these, the AVERAGE-OF-MAXIMA
method, returns the average of the variable values at which the maximum
truth value occurs.

For example, go back to our previous examples.  Using MAX-MIN inferencing
and AVERAGE-OF-MAXIMA defuzzification results in a crisp value of 8.4 for
z.  Using PRODUCT-SUM inferencing and CENTROID defuzzification results in
a crisp value of 5.6 for z, as follows.

Earlier on in the FAQ, we state that all variables (including z) take on
values in the range [0, 10].  To compute the centroid of the function f(x),
you divide the moment of the function by the area of the function.  To compute 
the moment of f(x), you compute the integral of x*f(x) dx, and to compute the
area of f(x), you compute the integral of f(x) dx.  In this case, we would
compute the area as integral from 0 to 10 of (0.32+0.036*z) dz, which is

    (0.32 * 10 + 0.018*100) =
    (3.2 + 1.8) =
    5.0

and the moment as the integral from 0 to 10 of (0.32*z+0.036*z*z) dz, which is

    (0.16 * 10 * 10 + 0.012 * 10 * 10 * 10) =
    (16 + 12) =
    28

Finally, the centroid is 28/5 or 5.6.

Note: Sometimes the composition and defuzzification processes are
combined, taking advantage of mathematical relationships that simplify
the process of computing the final output variable values.

The Mizumoto reference is probably "Improvement Methods of Fuzzy
Controls", in Proceedings of the 3rd IFSA Congress, pages 60-62, 1989.

================================================================
Go Back Up

Go To Previous

Go To Next