%PKLIBSVC Automatic radial basis SVM, using nu_libsvc and the Parzen kernel
%
%   W = PKLIBSVC(A,ALF)
%   W = A*PKLIBSVC([],A)
%   W = A*PKLIBSVC(A)
%
% INPUT
%   A   Dataset
%   ALF Parameter, default 1
%
% OUTPUT
%   W   Mapping: radial basis support vector classiifer
%
% DESCRIPTION
% This routine provides a radial basis support vector classifier based on
% NULIBSVC (which estimates NU using the leave-one-out 1NN error) and
% estimates the kernel width SIGMA by the the value found by PARZENC. The
% kernel width used is ALF*3*SQRT(2)*SIGMA. This is much faster than the
% gridsearch used by RBLIBSVC and performs about equal.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% DATASETS, MAPPINGS, LIBSVC, NULIBSVC, RBLIBSVC, PARZENC

% Copyright: R.P.W. Duin, r.p.w.duin@37steps.com

function w = pklibsvc(varargin)

  checktoolbox('libsvm');
	mapname = 'PK-LIBSVM';
  argin = shiftargin(varargin,'scalar');
  argin = setdefaults(argin,[],1);
  
  if mapping_task(argin,'definition')
    w = define_mapping(argin,'untrained',mapname);
    
  elseif mapping_task(argin,'training')			% Train a mapping.
  
    [a,alf] = deal(argin{:});

    if size(a,1) <= 1000
      [v,sig] = parzenc(a);
    else
      [v,sig] = parzenc(gendat(a,1000));
    end
    w = nulibsvc(a,proxm([],'r',alf*3*sig*sqrt(2)));
    
  end