Mercurial > hg > camir-aes2014

comparison toolboxes/FullBNT-1.0.7/nethelp3.3/rbftrain.htm @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
comparison
equal deleted inserted replaced
-1:000000000000 0:e9a9cd732c1e
1 <html>
2 <head>
3 <title>
4 Netlab Reference Manual rbftrain
5 </title>
6 </head>
7 <body>
8 <H1> rbftrain
9 </H1>
10 <h2>
11 Purpose
12 </h2>
13 Two stage training of RBF network.
14
15 <p><h2>
16 Description
17 </h2>
18 <CODE>net = rbftrain(net, options, x, t)</CODE> uses a
19 two stage training
20 algorithm to set the weights in the RBF model structure <CODE>net</CODE>.
21 Each row of <CODE>x</CODE> corresponds to one
22 input vector and each row of <CODE>t</CODE> contains the corresponding target vector.
23 The centres are determined by fitting a Gaussian mixture model
24 with circular covariances using the EM algorithm through a call to
25 <CODE>rbfsetbf</CODE>. (The mixture model is
26 initialised using a small number of iterations of the K-means algorithm.)
27 If the activation functions are Gaussians, then the basis function widths
28 are then set to the maximum inter-centre squared distance.
29
30 <p>For linear outputs,
31 the hidden to output
32 weights that give rise to the least squares solution
33 can then be determined using the pseudo-inverse. For neuroscale outputs,
34 the hidden to output weights are determined using the iterative shadow
35 targets algorithm.
36 Although this two stage
37 procedure may not give solutions with as low an error as using general
38 purpose non-linear optimisers, it is much faster.
39
40 <p>The options vector may have two rows: if this is the case, then the second row
41 is passed to <CODE>rbfsetbf</CODE>, which allows the user to specify a different
42 number iterations for RBF and GMM training.
43 The optional parameters to <CODE>rbftrain</CODE> have the following interpretations.
44
45 <p><CODE>options(1)</CODE> is set to 1 to display error values during EM training.
46
47 <p><CODE>options(2)</CODE> is a measure of the precision required for the value
48 of the weights <CODE>w</CODE> at the solution.
49
50 <p><CODE>options(3)</CODE> is a measure of the precision required of the objective
51 function at the solution. Both this and the previous condition must be
52 satisfied for termination.
53
54 <p><CODE>options(5)</CODE> is set to 1 if the basis functions parameters should remain
55 unchanged; default 0.
56
57 <p><CODE>options(6)</CODE> is set to 1 if the output layer weights should be should
58 set using PCA. This is only relevant for Neuroscale outputs; default 0.
59
60 <p><CODE>options(14)</CODE> is the maximum number of iterations for the shadow
61 targets algorithm;
62 default 100.
63
64 <p><h2>
65 Example
66 </h2>
67 The following example creates an RBF network and then trains it:
68 <PRE>
69
70 net = rbf(1, 4, 1, 'gaussian');
71 options(1, :) = foptions;
72 options(2, :) = foptions;
73 options(2, 14) = 10; % 10 iterations of EM
74 options(2, 5) = 1; % Check for covariance collapse in EM
75 net = rbftrain(net, options, x, t);
76 </PRE>
77
78
79 <p><h2>
80 See Also
81 </h2>
82 <CODE><a href="rbf.htm">rbf</a></CODE>, <CODE><a href="rbferr.htm">rbferr</a></CODE>, <CODE><a href="rbffwd.htm">rbffwd</a></CODE>, <CODE><a href="rbfgrad.htm">rbfgrad</a></CODE>, <CODE><a href="rbfpak.htm">rbfpak</a></CODE>, <CODE><a href="rbfunpak.htm">rbfunpak</a></CODE>, <CODE><a href="rbfsetbf.htm">rbfsetbf</a></CODE><hr>
83 <b>Pages:</b>
84 <a href="index.htm">Index</a>
85 <hr>
86 <p>Copyright (c) Ian T Nabney (1996-9)
87
88
89 </body>
90 </html>