Granular Neural Networks:

cracklegulleyAI and Robotics

Oct 19, 2013 (4 years and 2 months ago)

73 views

Granular Neural Networks:
Concepts and Development
Schemes

Mingli Song and Witold Pedrycz


IEEE Trans. On NNLS vol.24 2013

1

What is granular networks?


Introduced by W. Pedrycz and G. Vukovich in
Neurocomputing
,
2001



Traditional neural networks


numeric weights


single output



Granular neural networks


granular weights


interval output

2

Interval Operations


3

Architecture

4

k

j

m

1

i

1

n

X
1

X
i

X
n

O
j
=[o
j
-
, o
j
+
]

W
ji
=[W
ji
-
, W
ji
+
]

W
j
=[W
j
-
, W
j
+
]

Y=[y
-
, y
+
]

Granularity


ε
: values in unit interval


w
ji
-
=w
ji
-
ε
-
|w
ji
|, w
ji
+
=w
ji
+
ε
+
|w
ji
|



Allocation of information granularity:


C
1
: All weights use the same value of
ε
.
ε
-

=

ε
+
=

ε
/2


C
2
: All weights use the same value of
ε
.
ε
-
+

ε
+
=

ε


C
3
: Different granularity (
ε
i
) for different connections.
ε
i
-

=

ε
i+
=

ε
i
/2


C
4
: Different granularity (
ε
i
) for different connections.
ε
i
-
+

ε
i+
=

ε
i


C
5
: Randomly assigned.

5

Fitness Function


6

Single
-
objective PSO


7

Experiment

8