Information Technology Reference
In-Depth Information
to their parameters, such as neural networks, will be discussed. Finally, regu-
larization techniques, which aim at avoiding overfitting when training with a
small number of examples, will be discussed.
2.5.1 Training Models that are Linear with Respect to Their
Parameters: The Least Squares Method for Linear
Regression
We assume that the measurements of the quantity to be modeled can be
viewed as realizations of a random variable
Y
p
that is an a
ne function of
variables which have been selected in an earlier step:
Y
p
=
ζ
T
w
p
+
B
,where
ζ
is the vector of the variables of the model, of known dimension
q,
where
w
p
is the vector (non random but unknown) of the parameters of the model, and
where
B
is a random vector whose expectation value is zero. Therefore, the
regression function is linear with respect to the variables of the model
E
(
Y
p
)=
ζ
T
w
p
.
We want to design a model
g
(
ζ
,
w
)=
ζ
T
w
, given a set of
N
measurements of
the quantity of interest
y
p
,k
=1to
N
that are a set of realizations of the
random variable
Y
p
, and given a set of corresponding measurements of the
inputs
{
}
ζ
k
,k
=1to
N
{
}
.
2.5.1.1 Nonadaptive (Batch) Training of Models that are Linear
with Respect to Their Parameters
Because there is a wealth of textbooks on the subject (see for instance [Seber
1977; Antoniadis et al. 1992; Draper et al. 1998], no proof will be given in the
present section.
Minimizing the Least Squares Cost Function. The Normal Equations
The minimum of the following cost function is sought
J
(
w
)=
N
y
P
−
g
(
ζ
k
,
w
)
2
,
k
=1
with
g
(
ζ
,
w
)=
ζ
T
w
.
In such a model, the number of parameters
q
is equal to
the number of inputs
n
.
The matrix of observations is the matrix
Ξ
whose column
i
(
i
=1to
q
)
is the vector
ξ
i
whose components are the
N
measurements of the
i
th input:
therefore, it has
N
rows and
q
columns,
⎡
⎣
⎤
⎦
⎡
⎣
⎤
⎦
ζ
1
... ζ
1
(
ζ
1
)
T
...
...
...
(
ζ
N
)
T
n
... ... ...
... ... ...
... ... ...
ζ
1
... ζ
n
=
(
ζ
1
)
...
(
ζ
n
)
,
Ξ
=
=
Search WWH ::
Custom Search