Introduction to

Artificial Neural Networks

in Control
Andrew Paice 2009

1
Outline

2
Motivation

!
"
#
$
% &
"

$
"

3
Neurons

'
( ') !
* '* !
'+ "
!,
!

4
Artificial Neuron

"
"

5
Types of Neural Networks
'
,
-
, '
+ '

/! !
)
% &
/! !

6
Multi-Layer Perceptron Neural Network (MLPNN)

y1 y2 ym outputs

1
y=
1+ ex wo1

w11 w pm
( ' 1
wom w p1

z1 zp
" v o1
zj hidden layer

! 0 vop v np
v11
' 1

.
inputs
x1 x2 xn
*

7
Hopfield Neural Network

( '

+ !

2 "

!
8
Neural Network Learning (for MLPNNs)

1
'
y = f(x)
y NN(x,P(1), P(2), P(n))

" P(i)
!

"
{(xi,yi)}ni=1!
2 !
2 9
The Back-Propogation Algorithm
, " 1
y=
" 3 1 ex
2 !

y
= y (1 y )
x
!

NN ( n 1)
= y i (1 y i ) x j
Pij( n )

10
Backpropogation I
Step 1: Initialize weight to small random values.
Step 2: While stopping condition is false, do steps 3-10
Step 3: For each training pair do steps 4-9

Feedforward pass
Step 4: Each input unit receives the input signal xi and transmits this signals to all units in
the hidden layer
Step 5: Each input unit receives the input signal z j , j = 1,..., p sums its weighted input
signals
n
z inj = voj + xi vij
i =1

applying activation function Z j = f ( z inj ) and sends this to all units in the output layer
Step 6: Each output unit y k , k = 1,..., m sums its weighted input signals
p
y ink = wok + z j w jk
j =1

and applies its activation function to calculate the output signal Yk = f ( y ink )

11
Backpropogation II
Backward pass
Step 7: Each output unit y k , k = 1,..., m receives a target pattern corresponding to an input
pattern, error information term is calculated as
k = (t k y k ) f ( y ink )
Step 8: Each hidden unit z j , j = 1,..., p sums its delta inputs from units in the layer above
m
inj = j w jk
k =1
The error information term is calculated as
j = inj f ( z inj )

Updating Weight and Biases

Step 9: Each output unit y k , k = 1,..., m updates its bias and weights ( j = 0,..., p ) . The weight
correction term is given by
W jk = k z j
and the bias correction term is given by
Wok = k
Therefore, W jk (new) = W jk (old ) + W jk , Wok (new) = Wok (old ) + Wok
Each hidden unit z j , j = 1,..., p updates its bias and weights (i = 0,..., n) .
The weight correction term
Vij = j xi , Voj = j
Therefore, Vij (new) = Vij (old ) + Vij , Voj (new) = Voj (old ) + Voj
12
Step 10: Test the stopping condition, e.g. minimization of the errors, number of iterations, ...
Radial Basis Function Neural Network (RBFNN)
1
" 4

n
2
yj = wij ( x ci )
i =1

13
Effectiveness of NN Learning

2 "

%! ! )
1 / &

14
Application to Control
+ "

x+ = NN((x;u),P1)
y = NN((x;u),P2)

x+ = NN((Ax+Bu),P1)
y = NN(x,P2)

, "
yk+1 = NN((Y;U;wk),P1)
Y = (yk, yk-1, , yk-n)
U = (uk, uk-1, , uk-m)
15
Examples from CDC 2006
1(/

! "

4 )
" #$ $% !

0
5 ( !

16
Summary

%
&
"

3
6

% &
!

/ "
!

17
 )$. 1
+ 7

18