%RFISHERCC  Random fisher classifier combiner
%
%    W = RFISHERCC(A,N)
%    W = A*RFISHERCC(N)
%
% INPUT
%   A  prdataset to be used for training, M objects, C classes
%   N  number of base classifiers to be generated
%      default: M/(10*(C-1))
%
% OUTPUT
%   W  trained classifier
%
% DESCRIPTION
% This routine generates a a random set of N simple classifiers, based on
% the 1-NN rule using a single, randomly selected object per class. The
% confindences (shee KNNC) for the total training set A (in total N*(C-1)
% per object) are used to train a combiner using FISHERC.
%
% EXAMPLES
% a = gendatb;
% figure; scatterd(a); 
% plotc(a*rfishercc)
% 
% a = gendatm; 
% figure; scatterd(a); 
% plotc(a*rfishercc(2),'col')
%
% a = setprior(sonar,0); % make priors equal
% w1 = setname(rfishercc(10),'RFisher-10');
% w2 = setname(rfishercc(20),'RFisher-20');
% w3 = setname(rfishercc(40),'RFisher-40');
% randreset(1); % for reproducability
% e = cleval(a,{w1,w2,w3},[5,10,20,40,80],10);
% plote(e);
%
% SEE ALSO(<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, KNNC, FISHERC

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

function w = rfishercc(varargin)

argin = shiftargin(varargin,'scalar');
argin = setdefaults(argin,[],[]);
if mapping_task(argin,'definition')
  w = define_mapping(argin,'untrained');
  
elseif mapping_task(argin,'training')
  [a,n] = deal(argin{:});
  isdataset(a);
  islabtype(a,'crisp');
  [m,k,c] = getsize(a);
  
  % default number of classifiers, <= 500
  if isempty(n)
    n = min(ceil(m/(10*(c-1))),500);
  end
  
  u = a*repmat({gendat([],ones(1,c))*knnc([],1)},1,n);  
  u = u*featsel(1:c-1);
  w = a*(stacked(u)*fisherc);
  
else
  error('Wrong input')
end