87-21Applied Intelligence Processing
4. Use a recursive algorithm starting at the output layer and adjust the weights backward by
w t w t x
ij ij j i
( ) ( )+ = +
′
1
ηδ
where
w
ij
(t) is the weight from an input to node j at time t
′
i
is either the output of node or an input
η
i
s
a g
ain
t
erm
(
0
.0 < η < 1.0)
δ
j
is an error term for node j
For
t
he
o
utput
l
ayer
l = L, t
he
e
rror
i
s
c
alculated:
δ
j
L
j
L
j j
g h d y=
′
−[ ]
where
L
j
L
is the net input to the th unit in the layer
g′
i
s
t
he
d
erivative
o
f
t
he
a
ctivation
f
unction
g
F
or hidden layers l = (L − 1), …, 1, the error is calculated:
δ δ
j
L
j
l
ij
l
j
l
l
g h w =
′
+ +
1 1
where
l
j
l
is the net input to the th unit in the th layer
g′
i
s
t
he
d
erivative
o
f
t
he
a
ctivation
f
unction
g
5. Re
turn
to Ste
p
2 an
d
re ...