Information Technology Reference
In-Depth Information
For practical applications, the designer must understand the basic concepts
in order to obtain reliable results, but he must also use appropriate tools (or
build his own, which may be a lengthy process).
At present, available development tools fall into two categories:
•
neural network toolboxes within general-purpose engineering software;
typically, Matlab releases a toolbox that allows easy training of feedforward
neural networks; the programming effort is very small for classical func-
tions, but it may become important for the implementation of elements
of methodology that are not specifically “neural” (leverage computation,
input selection), or for recurrent neural networks;
•
specific development tools that include a complete development method-
ology, requiring no programming effort; typical is the NeuroOne
6
package;
such tools do not allow for the flexibility of personal programming, but
they provide reliable results in a short time.
Some academic software packages are available freely on the Web. They are
excellent for didactic purposes, but they may not stand up to the quality
requirements of industrial applications.
Therefore, the model designer, whether in academy or industry, must
choose his tools considering the time constraints, the development policy
within the company, the size of the applications, etc. The best solution con-
sists in having both types of tools available. Anyway, however powerful and
user-friendly the programming tools, a good understanding of the basic con-
cepts and methods, and the application of a principled methodology, are the
keys to the development of successful applications.
2.10 Additional Material
This section is devoted to additional definitions, proofs, algorithms, which can
be skipped on first reading.
2.10.1 Confidence Intervals: Design and Example
2.10.1.1 Design
In order to estimate a confidence interval for a random variable
Y
, one seeks
a random variable
Z
, function of
Y
, whose distribution
p
Z
(
z
)isknownand
independent of
Y
. Since the distribution
p
Z
(
z
) is known and tabulated, the
equation Pr(
z
1
<z<z
2
)=
z
2
z
1
α
can be solved easily: one just
has to compute the value of
z
1
such that Pr(
z<z
1
)=
α/
2, and the value of
z
2
such that Pr(
z>z
2
)=
α/
2. When
z
1
and
z
2
are found, function
Z
(
Y
)is
inverted in order to find the values of
a
et
b
such that Pr(
a<y<b
)=1
p
Z
(
z
)d
z
=1
−
−
α
.
6
By NETRAL S.A.; several illustrations and applications described in Chaps. 1
and 2 were developed with that software.
Search WWH ::
Custom Search