www.pudn.com > Classification_toolbox.part01.rar > Balanced_Winnow.m

function [test_targets, a_plus, a_minus] = Balanced_Winnow(train_patterns, train_targets, test_patterns, params) 
 
% Classify using the balanced Winnow algorithm 
% Inputs: 
% 	training_patterns   - Train patterns 
%	training_targets	- Train targets 
%   test_patterns       - Test  patterns 
%	params		        - [Num iter, Alpha, Convergence rate] 
% 
% Outputs 
%	test_targets        - Predicted targets 
%   a_plus	            - The positive weight vector 
%   a_minus	            - The negative weight vector 
 
[c, r]		    		  = size(train_patterns); 
[Max_iter, alpha, eta]    = process_params(params); 
y 			              = [train_patterns ; ones(1,r)]; 
z        			      = train_targets; 
 
%Initial weights 
a_plus	        = sum(y')'; 
a_minus	        = -sum(y')'; 
iter  	        = 0; 
 
while (iter < Max_iter) 
   iter = iter + 1; 
    
   for k = 1:r, 
      if (sign(a_plus'*y(:,k) - a_minus'*y(:,k)) ~= sign(z(k)-.5)), 
         if (z(k) == 1), 
            a_plus	= alpha.^y(:,k).*a_plus; 
            a_minus	= alpha.^-y(:,k).*a_minus; 
         else 
            a_plus	= alpha.^-y(:,k).*a_plus; 
            a_minus	= alpha.^y(:,k).*a_minus; 
         end 
      end 
   end 
    
end 
 
a = (a_plus + a_minus)/2; 
 
%Classify the test patterns 
test_targets = a' * [test_patterns; ones(1,size(test_patterns,2))]; 
 
if (length(unique(train_targets))== 2) 
    test_targets = test_targets < 0.5; 
end