| RSRBF | |
Redused Set Method for RBF kernel expansion.
Synopsis:
red_model = rsrbf(model)
red_model = rsrbf(model,options)
Description:
red_model = rsrbf(model) searchs for a kernel expansion
with nsv vectors which best approximates the input
expansion given in model [Schol98a]. The Radial Basis
kernel (RBF) is assumed (see 'help kernel').
red_model = rsrbf(model,options) allows to specify the
control paramaters.
Input:
model [struct] Kernel expansion:
.Alpha [nsv x 1] Weight vector.
.sv.X [dim x nsv] Vectors defining the expansion.
.options.ker [string] Must be equal to 'rbf'.
.options.arg [1x1] Kernel argument (see 'help kernel').
options [struct] Control parameters:
.nsv [1x1] Desired number of vectors in the reduced
expansion (default round(length(model.Alpha)/2)).
.eps [1x1] Desier limit on the norm of difference between
the original normal vector and the reduced the normal
vector in the feature space. The algorithm is stopped
when a lower difference is achived (default 1e-6).
.preimage [string] Function called to solve the RBF pre-image
problem (default 'rbfpreimg');
.verb [1x1] If 1 then progress info is display (default 0).
Output:
red_model [struct] Reduced kernel expansion.
Example:
trn = load('riply_trn');
model = smo(trn,struct('ker','rbf','arg',1,'C',10));
red_model = rsrbf(model,struct('nsv',10));
figure; ppatterns(trn);
h1 = pboundary(model,struct('line_style','r'));
h2 = pboundary(red_model,struct('line_style','b'));
legend([h1(1) h2(1)],'Original SVM','Reduced SVM');
See also
RBFPREIMG.
About: Statistical Pattern Recognition Toolbox
(C) 1999-2003, Written by Vojtech Franc and Vaclav Hlavac
Czech Technical University Prague
Faculty of Electrical Engineering
Center for Machine Perception
Modifications:
11-oct-2004, VF, knorm.m used
21-sep-2004, VF
10-jun-2004, VF
02-dec-2003, VF