Information Technology Reference
In-Depth Information
z
−
1
z
−
1
z
−
1
z
−
1
x
(
τ
)
Bias
y
(
τ
)
u
(
τ
)
Fig. 6.12 A fully recurrent one hidden layer neural network. The notation is ex-
plained in the text.
6.2.1.3
Application of ZED to RTRL
The application of the ZED in this case cannot be based, as previously done,
using a static set of
n
error values. Instead, one has to consider
L
previous
values of the error in order to build a dynamic approximation of the error
density. This will use a time sliding window that will allow the definition of
the density in an online formulation of the learning problem.
For a training set of size
n
, the error r.v. e(
i
) represents the differ-
ence between the desired output vector and the actual output, for a given