kperceptr.m

支持向量机工具箱

function [Alpha,bias,sol,t,nsv]=...
    kperceptr(data,labels,ker,arg,tmax)
% KPERCEPTR Kernel Perceptron algorithm.
%
% [Alpha,bias,sol,t,nsv]=kperceptr(data,labels,ker,arg,tmax)
%
% Input:
%  data [dim x n] input patterns; dim is dimension and
%    n is number of patterns.
%  labels [1 x n] labels of patterns; 1 denotes the 1st class
%    2 denotes the 2nd class.
%  ker [string] kernel identifier (see 'help kernel').
%  arg [...] argumetn of given kernel.
%  tmax [int] maximal number of iterations.
%
% Output:
%  Alpha [1 x n] positive linear pattern multipliers.
%  bias [real] bias of the found decision function.
%  sol [int] 1 - Perceptron converged to the solution (zero
%   training classification error), 0 - Perceptron has not
%   converged in tmax iterations.
%  t [int] number of iterations.
%  nsv [int] number of non-zero multipliers Alpha.
% 
% See also PERCEPTR, SVM, SVMCLASS, PSVM.
%

% Statistical Pattern Recognition Toolbox, Vojtech Franc, Vaclav Hlavac
% (c) Czech Technical University Prague, http://cmp.felk.cvut.cz
%
% Modifications: 
%  21-Nov-2001, V. Franc

if nargin < 4,
  error('Not enough input arguments.');
end

if nargin <5,
  tmax = inf;
end

% get dimension and number of training patterns
[dim, n ] = size( data );

% precompute kernel function
K = kmatrix( data, ker, arg );

% make labels to be 1 and -1
Y = itosgn( labels );

% inicialize multiliers Alpha and bias
Alpha = zeros( 1, n );
bias = 0;

sol = 0;
t = 0;

while t < tmax & sol == 0,
  
  t = t + 1;
  
  % dot products < x, w >
  proj = Y'.*(K*(Y.*Alpha)'+bias);
  
  % find misclassified patter
  [dot_prod, inx ] = min( proj );
  
  % chacks whether the adaptation is needed
  if dot_prod <= 0,
    Alpha( inx ) = Alpha( inx ) + 1;
    bias = Y(inx) * bias;
  else
    sol = 1;
  end
end

% compute number of nonzero Alphas
nsv = length( find( Alpha ));

return;