|
Daniel@0
|
1 <html>
|
|
Daniel@0
|
2 <head>
|
|
Daniel@0
|
3 <title>
|
|
Daniel@0
|
4 Netlab Reference Manual glmtrain
|
|
Daniel@0
|
5 </title>
|
|
Daniel@0
|
6 </head>
|
|
Daniel@0
|
7 <body>
|
|
Daniel@0
|
8 <H1> glmtrain
|
|
Daniel@0
|
9 </H1>
|
|
Daniel@0
|
10 <h2>
|
|
Daniel@0
|
11 Purpose
|
|
Daniel@0
|
12 </h2>
|
|
Daniel@0
|
13 Specialised training of generalized linear model
|
|
Daniel@0
|
14
|
|
Daniel@0
|
15 <p><h2>
|
|
Daniel@0
|
16 Description
|
|
Daniel@0
|
17 </h2>
|
|
Daniel@0
|
18 <CODE>net = glmtrain(net, options, x, t)</CODE> uses
|
|
Daniel@0
|
19 the iterative reweighted least squares (IRLS)
|
|
Daniel@0
|
20 algorithm to set the weights in the generalized linear model structure
|
|
Daniel@0
|
21 <CODE>net</CODE>. This is a more efficient alternative to using <CODE>glmerr</CODE>
|
|
Daniel@0
|
22 and <CODE>glmgrad</CODE> and a non-linear optimisation routine through
|
|
Daniel@0
|
23 <CODE>netopt</CODE>.
|
|
Daniel@0
|
24 Note that for linear outputs, a single pass through the
|
|
Daniel@0
|
25 algorithm is all that is required, since the error function is quadratic in
|
|
Daniel@0
|
26 the weights. The algorithm also handles scalar <CODE>alpha</CODE> and <CODE>beta</CODE>
|
|
Daniel@0
|
27 terms. If you want to use more complicated priors, you should use
|
|
Daniel@0
|
28 general-purpose non-linear optimisation algorithms.
|
|
Daniel@0
|
29
|
|
Daniel@0
|
30 <p>For logistic and softmax outputs, general priors can be handled, although
|
|
Daniel@0
|
31 this requires the pseudo-inverse of the Hessian, giving up the better
|
|
Daniel@0
|
32 conditioning and some of the speed advantage of the normal form equations.
|
|
Daniel@0
|
33
|
|
Daniel@0
|
34 <p>The error function value at the final set of weights is returned
|
|
Daniel@0
|
35 in <CODE>options(8)</CODE>.
|
|
Daniel@0
|
36 Each row of <CODE>x</CODE> corresponds to one
|
|
Daniel@0
|
37 input vector and each row of <CODE>t</CODE> corresponds to one target vector.
|
|
Daniel@0
|
38
|
|
Daniel@0
|
39 <p>The optional parameters have the following interpretations.
|
|
Daniel@0
|
40
|
|
Daniel@0
|
41 <p><CODE>options(1)</CODE> is set to 1 to display error values during training.
|
|
Daniel@0
|
42 If <CODE>options(1)</CODE> is set to 0,
|
|
Daniel@0
|
43 then only warning messages are displayed. If <CODE>options(1)</CODE> is -1,
|
|
Daniel@0
|
44 then nothing is displayed.
|
|
Daniel@0
|
45
|
|
Daniel@0
|
46 <p><CODE>options(2)</CODE> is a measure of the precision required for the value
|
|
Daniel@0
|
47 of the weights <CODE>w</CODE> at the solution.
|
|
Daniel@0
|
48
|
|
Daniel@0
|
49 <p><CODE>options(3)</CODE> is a measure of the precision required of the objective
|
|
Daniel@0
|
50 function at the solution. Both this and the previous condition must be
|
|
Daniel@0
|
51 satisfied for termination.
|
|
Daniel@0
|
52
|
|
Daniel@0
|
53 <p><CODE>options(5)</CODE> is set to 1 if an approximation to the Hessian (which assumes
|
|
Daniel@0
|
54 that all outputs are independent) is used for softmax outputs. With the default
|
|
Daniel@0
|
55 value of 0 the exact Hessian (which is more expensive to compute) is used.
|
|
Daniel@0
|
56
|
|
Daniel@0
|
57 <p><CODE>options(14)</CODE> is the maximum number of iterations for the IRLS algorithm;
|
|
Daniel@0
|
58 default 100.
|
|
Daniel@0
|
59
|
|
Daniel@0
|
60 <p><h2>
|
|
Daniel@0
|
61 See Also
|
|
Daniel@0
|
62 </h2>
|
|
Daniel@0
|
63 <CODE><a href="glm.htm">glm</a></CODE>, <CODE><a href="glmerr.htm">glmerr</a></CODE>, <CODE><a href="glmgrad.htm">glmgrad</a></CODE><hr>
|
|
Daniel@0
|
64 <b>Pages:</b>
|
|
Daniel@0
|
65 <a href="index.htm">Index</a>
|
|
Daniel@0
|
66 <hr>
|
|
Daniel@0
|
67 <p>Copyright (c) Ian T Nabney (1996-9)
|
|
Daniel@0
|
68
|
|
Daniel@0
|
69
|
|
Daniel@0
|
70 </body>
|
|
Daniel@0
|
71 </html> |
