GGRADANDER |
Gradient method to solve the Generalized Anderson's task.
Synopsis:
model = ggradandr( distrib)
model = ggradandr( distrib, options)
model = ggradandr( distrib, options, init_model)
Description:
This function is an implementation of the algorithm
using the generalized gradient optimization to solve
the the Generalized Anderson's task [SH10].
The goal of the GAT is find the binary linear classification
rule (g(x)=sgn(W'*x+b) with minimal probability of
misclassification. The conditional probabilities are known to
be Gaussians their paramaters belong to a given set of parameters.
The true parameters are not known. The linear rule which
guarantes the minimimal classification error for the worst
possible case (the worst configuration of Gaussains) is
sought for.
Input:
distrib [struct] Binary labeled Gaussian distributions:
.Mean [dim x ncomp] Mean vectors.
.Cov [dim x dim x ncomp] Covariance matrices.
.y [1 x ncomp] labels of Gaussians (1 or 2).
options [struct] Defines stopping condition:
.tmax [1x1] Maximal number of iterations (default 1e4 ).
.eps [1x1] Minimal change in the optimised criterion (default 0).
init_model [struct] Initial model:
.W, .b, .t see below.
Output:
model [struct] Linear classifier:
.W [dim x 1] Normal vector of the found hypeprlane W'*x + b = 0.
.b [1x1] Bias of the hyperplane.
.r [1x1] Mahalanobis distance for the cloasest Gaussian.
.err [1x1] Probability of misclassification.
.t [1x1] Number of iterations.
.exitflag [1x1] 0 ... maximal number of iterations exceeded.
1 ... condition abs( r - old_r) < eps fulfilled.
Example:
distrib = load('mars');
model = ggradandr( distrib );
figure; pandr( model, distrib );
See also
ANDRORIG, EANDERS, GANDERS, ANDRERR, LINCLASS.
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:
17-sep-2003, VF