Sie sind auf Seite 1von 19

FLANN

Prof. G. Panda
IIT Bhubaneswar
Functional Link Artificial Neural Network
(FLANN)
Prof. Ganapati Panda, FNAE, FNASc.
Deputy Director
Indian Institute Of Technology Bhubaneswar
FLANN
Prof. G. Panda
IIT Bhubaneswar
The functional link ANN or Pao-network originally
proposed by Pao.
Single layer ANN structure
Need of hidden layer is removed
Capable of forming arbitrarily complex decision
regions
Offers less computational complexity
Pao, Y. H. (1989). Adaptive pattern recognition & neural networks. Reading,
MA: Addison-Wesley
FLANN
Prof. G. Panda
IIT Bhubaneswar
Structure of FLANN
e(k)
X
X
X
X
X
X
FE
1
FE
2
FE
3
X
1
(1)
X
5
(1)
X
2
(1)
X
2
(5)
X
3
(1)
X
3
(5)
W
1
(k)
W
5
(k)
W
6
(k)
W
10
(k)
W
11
(k)
W
15
(k)
Adaptive Algorithm
y(k)
Functional
Expansion
x
1
(k)
x
1
(k-1)
x
1
(k-2)
d(k)
E
E
+
-
FLANN
Prof. G. Panda
IIT Bhubaneswar
Functional expansions
Trigonometric
Polynomial
Exponential
Chebyshev
FLANN
Prof. G. Panda
IIT Bhubaneswar
x
1
(k)
x
1
(k)








sin(pi*x
1
(k))
cos(pi*x
1
(k))
sin(2*pi*x
1
(k))
cos(2*pi*x
1
(k))
Trigonometric Expansion
FLANN
Prof. G. Panda
IIT Bhubaneswar
E
0
x
1
x
2
1
x
2 1
x x
2
x
2
2
x
1
x
2
x
Polynomial Expansion
FLANN
Prof. G. Panda
IIT Bhubaneswar
T
n+1
(x) = 2x T
n
(x) - T
n-1
(x)

The First Few Chebyschev Polynomials
T
o
(x) = 1
T
1
(x) = x
T
2
(x) = 2x
2
1
T
3
(x) = 4x
3
3x
T
4
(x) = 8x
4
8x
2
+ 1
T
5
(x) = 16x
5
20x
3
+ 5x

x
T
1
(x)
T
2
(x)
T
3
(k)
T
4
(x)
T
5
(x)
Chebyschev Expansion
FLANN
Prof. G. Panda
IIT Bhubaneswar
Learning rule
W
t
(k+1)=W
t
(k) + mu * e(k)*X(k)

Where
mu=convergence coefficient
e(k) = error at kth instant
X(k) = expanded input vector at kth instant
FLANN
Prof. G. Panda
IIT Bhubaneswar
FLANN with sigmoid
e(k)
d(k)
X
X
X
X
X
X
FE
1
FE
2
FE
3
X
1
(1)
X
5
(1)
X
2
(1)
X
2
(5)
X
3
(1)
X
3
(5)
W
1
(k)
W
5
(k)
W
6
(k)
W
10
(k)
W
11
(k)
W
15
(k)
Adaptive Algorithm
y(k)
Functional
Expansion
x
1
(k)
x
1
(k-1)
x
1
(k-2)
E
E
+
-
tanh
bias
X
FLANN
Prof. G. Panda
IIT Bhubaneswar
Delta Learning
The error is given by
) ( ) ( ) ( k y k d k e =
Then the delta is calculated as
) ( * ) ) ( 1 ( ) (
2
k e k y k = o
Now the weight update equation becomes
) 1 ( ) ( ) ( ) 1 ( A + A + = + k k k W k W
T
k
X k k )] [( ) ( ) ( o = A
where
FLANN
Prof. G. Panda
IIT Bhubaneswar
Epoch based learning
) (
1
) (
1
k w
N
i w
N
k

=
A = A
) ( ) ( ) 1 ( i w i w i w A + = +
b b b
w i w i w A + = + ) ( ) 1 (
Application of all patterns constitutes one
experiment. At the end of each experiment N
sets of () are obtained . Then the average
change of weight is computed as


The weights are updated as


Bias weight is updated as
FLANN
Prof. G. Panda
IIT Bhubaneswar
Function Approximation
x x x x f 4 . 0 3 . 0 ) (
2 3
1
+ =
) 5 sin( 1 . 0 ) 3 sin( 3 . 0 ) sin( 6 . 0 ) (
2
x x x x f t t t + + =
The input pattern is expanded using trigonometric
expansion
Fifteen input nodes including a bias input are used
The nonlinearity associated is tanh( ) function.
The convergence coefficient is set to 0.1
Training using an uniformly distributed random signal over
the interval [-1,1].

FLANN
Prof. G. Panda
IIT Bhubaneswar
Function approximation

> +
s
=
250
25
2
sin 2 . 0
250
2
sin 8 . 0
250
250
2
sin
) (
k for
k k
k for
k
k x
t t
t
During testing the input to the identified model is
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Discrete Time
O
u
t
p
u
t
s
Plant
Model
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Discrete Time
O
u
t
p
u
t
s
Plant
Model
Example-1
Example-2
FLANN
Prof. G. Panda
IIT Bhubaneswar
E
System NL
Model
Adaptive
algorithm
x(k)
Random
noise
y(k)
y(k)
e(k)
E
+
-
+
System Identification
FLANN
Prof. G. Panda
IIT Bhubaneswar
System Identification
Input signal : uniformly distributed random signal [-.5,.5].
SNR : 30dB
H(z) : 0.260 + 0.930z
1
+ 0.260z
2
The learning parameter : 0.02
Iteration : 2000
Averaging : 50 times
J. C. Patra, R. N. Pal, B. N. Chatterjee, G. Panda, Identification Of Nonlinear
Dynamic Systems Using Functional Link Artificial Neural Network, IEEE Trans.
Systems, Man, and Cybernetics, 29(2), pp: 254-262,1999. [Cited 151 times : Dec 11]
FLANN
Prof. G. Panda
IIT Bhubaneswar
System Identification
0 500 1000 1500 2000
-35
-30
-25
-20
-15
-10
-5
0
No of iteration

M
S
E

i
n

d
B

MLP
CFLANN
FLANN
NL : y=x
0 500 1000 1500 2000
-35
-30
-25
-20
-15
-10
-5
0
No of iteration

M
S
E

i
n

d
B

MLP
CFLANN
FLANN
NL : tanh(.)
FLANN
Prof. G. Panda
IIT Bhubaneswar
Conclusions
Simple structure

Less computational complexity

Faster learning
FLANN
Prof. G. Panda
IIT Bhubaneswar
Latest References
1. Nithin V. George and G. Panda, On the development of adaptive hybrid active
noise control system for effective mitigation of nonlinear noise, Signal Processing,
92(2), pp: 509-516, 2012.
2. R. Majhi, B. Majhi and G. Panda, The development and performance evaluation of
neural network classifiers for Indian internet shoppers, Expert Systems with
Applications, 39(2), pp: 2112-2118, 2012.
3. Babita Majhi and G. Panda, Robust identification of nonlinear complex systems
using low complexity ANN and particle swarm optimization technique, Expert
Systems with Applications, 38(1), pp: 321-333, 2011.
4. S. K. Mishra, G. Panda and D. P. Das, A Novel Method of Extending the Linearity
Range of LVDT Using Artificial Neural Network, IEEE Trans. on Instrumentation and
Measurement, 59(4), pp: 947-953, 2010.
FLANN
Prof. G. Panda
IIT Bhubaneswar






THANK YOU

Das könnte Ihnen auch gefallen