Information Technology Reference
In-Depth Information
The
segment detector
is used to identify the intersecting segments. Registers A1
and B1 store the upper end-points while the registers A2 and B2 store the lower end-
points. If an intersection is found, the contents of the registers are transferred to
p
-
∆X
calculator
. Otherwise the contents of A1 and B1 are transferred to A2 and B2 respec-
tively and A1 and B1 get the upper end-points of the next segments from M
3
and M
4
respectively.
p
and
∆X
are used in the search algorithm to find the intersection point.
2.3
Rule Unit
The rule unit is used for rule activation. There are 8 rule units for parallel inferencing
of the fuzzy rules. The architectural design of the rule unit is shown in Fig. 6. Each
unit has 3 memories M
5
, M
6
and M
7
. M
5
and M
6
store the antecedents of the rules and
the consequents are stored in M
7
.
Fig. 6.
Rule Unit
Usually a group of fuzzy rules have a set of common antecedents. The set of com-
mon antecedents for the group of rules are stored in M
5
and the varying antecedents
are stored in M
6
. This leads to savings in memory space, computation time and power
dissipation. The total number of rules is 256. M
5
contains 8 words, each 24 bits wide.
The first 16 bits are used to indicate the fuzzy variables which are common over a
group of rules. The next four bits show the linguistic element and the last 4 bits give
the number of rules in that group. Each group on an average has 4 rules. M
6
has 32
words each 20 bits wide. The first 16 bits indicate the uncommon fuzzy variables and
the next 4 bits show the linguistic element.
The consequent of each rule is a membership function. Similar to the antecedent
membership functions, the output membership functions are approximated using
piece-wise linear segments. Each membership function has 7 linear segments. The
abscissas of the endpoints are stored in M
7
, with the ordinates fixed and known. Thus
each consequent requires 8 memory words. M
7
has therefore 32 * 8 = 256 words, each
being 16 bits wide.
The process of rule activation involves minimization of the truth degrees of the ante-
cedents. If the degree of truth for some antecedent is null, then that rule becomes inac-
tive and is skipped. A word from memory M
5
is read and each antecedent is checked for
its degree of truth in the
antecedent tester
. This is done by reading the corresponding
indicator variable
a
ik
from the antecedent memory. If the indicator variable is zero, then
the degree of truth for that antecedent is zero and as a result all the rules in that group
are inactive. The control then passes onto the next word of M
5
(next group of rules). If
the indicator variable is 1, then that antecedent has a positive degree of truth. The truth
value is read from the
ʱ
-memory and sent to the
minima calculator
.