wolffd@0: function codebook = lvq3(codebook,data,rlen,alpha,win,epsilon)
wolffd@0: 
wolffd@0: %LVQ3 trains codebook with LVQ3 -algorithm
wolffd@0: %
wolffd@0: % sM = lvq3(sM,D,rlen,alpha,win,epsilon)
wolffd@0: %
wolffd@0: %   sM = lvq3(sM,sD,50*length(sM.codebook),0.05,0.2,0.3);
wolffd@0: %
wolffd@0: %  Input and output arguments: 
wolffd@0: %   sM      (struct) map struct, the class information must be 
wolffd@0: %                    present on the first column of .labels field
wolffd@0: %   D       (struct) data struct, the class information must
wolffd@0: %                    be present on the first column of .labels field
wolffd@0: %   rlen    (scalar) running length
wolffd@0: %   alpha   (scalar) learning parameter, e.g. 0.05
wolffd@0: %   win     (scalar) window width parameter, e.g. 0.25
wolffd@0: %   epsilon (scalar) relative learning parameter, e.g. 0.3
wolffd@0: %
wolffd@0: %   sM      (struct) map struct, the trained codebook
wolffd@0: %
wolffd@0: % NOTE: does not take mask into account.
wolffd@0: %
wolffd@0: % For more help, try 'type lvq3', or check out online documentation.
wolffd@0: % See also LVQ1, SOM_SUPERVISED, SOM_SEQTRAIN.
wolffd@0: 
wolffd@0: %%%%%%%%%%%%% DETAILED DESCRIPTION %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: %
wolffd@0: % lvq3
wolffd@0: %
wolffd@0: % PURPOSE
wolffd@0: %
wolffd@0: % Trains codebook with the LVQ3 -algorithm (described below).
wolffd@0: %
wolffd@0: % SYNTAX
wolffd@0: %
wolffd@0: % sM = lvq3(sM, data, rlen, alpha, win, epsilon)
wolffd@0: %
wolffd@0: % DESCRIPTION
wolffd@0: %
wolffd@0: % Trains codebook with the LVQ3 -algorithm. Codebook contains a number
wolffd@0: % of vectors (mi, i=1,2,...,n) and so does data (vectors xj, j=1,2,...k).
wolffd@0: % Both vector sets are classified: vectors may have a class (classes are
wolffd@0: % set to data- or map -structure's 'labels' -field. For each xj the two 
wolffd@0: % closest codebookvectors mc1 and mc2 are searched (euclidean distances
wolffd@0: % d1 and d2). xj must fall into the zone of window. That happens if:
wolffd@0: %
wolffd@0: %    min(d1/d2, d2/d1) > s, where s = (1-win) / (1+win).
wolffd@0: %
wolffd@0: % If xj belongs to the same class of one of the mc1 and mc1, codebook
wolffd@0: % is updated as follows (let mc1 belong to the same class as xj):
wolffd@0: %    mc1(t+1) = mc1(t) + alpha * (xj(t) - mc1(t))
wolffd@0: %    mc2(t+1) = mc2(t) - alpha * (xj(t) - mc2(t))
wolffd@0: % If both mc1 and mc2 belong to the same class as xj, codebook is
wolffd@0: % updated as follows:
wolffd@0: %    mc1(t+1) = mc1(t) + epsilon * alpha * (xj(t) - mc1(t))
wolffd@0: %    mc2(t+1) = mc2(t) + epsilon * alpha * (xj(t) - mc2(t))
wolffd@0: % Otherwise updating is not performed.
wolffd@0: %
wolffd@0: % Argument 'rlen' tells how many times training -sequence is performed.
wolffd@0: %
wolffd@0: % Argument 'alpha' is recommended to be smaller than 0.1 and argument
wolffd@0: % 'epsilon' should be between 0.1 and 0.5.
wolffd@0: %
wolffd@0: % NOTE: does not take mask into account.
wolffd@0: %
wolffd@0: % REFERENCES
wolffd@0: %
wolffd@0: % Kohonen, T., "Self-Organizing Map", 2nd ed., Springer-Verlag, 
wolffd@0: %    Berlin, 1995, pp. 181-182.
wolffd@0: %
wolffd@0: % See also LVQ_PAK from http://www.cis.hut.fi/research/som_lvq_pak.shtml
wolffd@0: % 
wolffd@0: % REQUIRED INPUT ARGUMENTS
wolffd@0: %
wolffd@0: %  sM                The data to be trained.
wolffd@0: %          (struct)  A map struct.
wolffd@0: %
wolffd@0: %  data              The data to use in training.
wolffd@0: %          (struct)  A data struct.
wolffd@0: %
wolffd@0: %  rlen    (integer) Running length of LVQ3 -algorithm.
wolffd@0: %                    
wolffd@0: %  alpha   (float)   Learning rate used in training, e.g. 0.05
wolffd@0: %
wolffd@0: %  win     (float)   Window length, e.g. 0.25
wolffd@0: %  
wolffd@0: %  epsilon (float)   Relative learning parameter, e.g. 0.3
wolffd@0: %
wolffd@0: % OUTPUT ARGUMENTS
wolffd@0: %
wolffd@0: %  sM          Trained data.
wolffd@0: %          (struct)  A map struct.
wolffd@0: %
wolffd@0: % EXAMPLE
wolffd@0: %
wolffd@0: %   lab = unique(sD.labels(:,1));         % different classes
wolffd@0: %   mu = length(lab)*5;                   % 5 prototypes for each    
wolffd@0: %   sM = som_randinit(sD,'msize',[mu 1]); % initial prototypes
wolffd@0: %   sM.labels = [lab;lab;lab;lab;lab];    % their classes
wolffd@0: %   sM = lvq1(sM,sD,50*mu,0.05);          % use LVQ1 to adjust
wolffd@0: %                                         % the prototypes      
wolffd@0: %   sM = lvq3(sM,sD,50*mu,0.05,0.2,0.3);  % then use LVQ3 
wolffd@0: % 
wolffd@0: % SEE ALSO
wolffd@0: % 
wolffd@0: %  lvq1             Use LVQ1 algorithm for training.
wolffd@0: %  som_supervised   Train SOM using supervised training.
wolffd@0: %  som_seqtrain     Train SOM with sequential algorithm.
wolffd@0: 
wolffd@0: % Contributed to SOM Toolbox vs2, February 2nd, 2000 by Juha Parhankangas
wolffd@0: % Copyright (c) by Juha Parhankangas
wolffd@0: % http://www.cis.hut.fi/projects/somtoolbox/
wolffd@0: 
wolffd@0: % Juha Parhankangas 310100 juuso 020200
wolffd@0: 
wolffd@0: NOTFOUND = 1;
wolffd@0: 
wolffd@0: cod = codebook.codebook;
wolffd@0: dat = data.data;
wolffd@0: 
wolffd@0: c_class = codebook.labels(:,1);
wolffd@0: d_class = data.labels(:,1);
wolffd@0: 
wolffd@0: s = (1-win)/(1+win);
wolffd@0: 
wolffd@0: x = size(dat,1);
wolffd@0: y = size(cod,2);
wolffd@0: 
wolffd@0: c_class=class2num(c_class);
wolffd@0: d_class=class2num(d_class);
wolffd@0: 
wolffd@0: ONES=ones(size(cod,1),1);
wolffd@0: 
wolffd@0: for t=1:rlen
wolffd@0:   fprintf('\rTraining round: %d/%d',t,rlen);
wolffd@0:   tmp = NaN*ones(x,y);
wolffd@0:  
wolffd@0:   for j=1:x
wolffd@0:     flag = 0;
wolffd@0:     mj = 0;
wolffd@0:     mi = 0;
wolffd@0:     no_NaN=find(~isnan(dat(j,:)));
wolffd@0:     di=sqrt(sum([cod(:,no_NaN) - ONES*dat(j,no_NaN)].^2,2));
wolffd@0:     [foo, ind1] = min(di);
wolffd@0:     di(ind1)=Inf;
wolffd@0:     [foo,ind2] =  min(di);    
wolffd@0:   
wolffd@0:     %ind2=ind2+1;
wolffd@0: 
wolffd@0:     if d_class(j) & d_class(j)==c_class(ind1)
wolffd@0:       mj = ind1;
wolffd@0:       mi = ind2;
wolffd@0:       if d_class(j)==c_class(ind2)
wolffd@0:         flag = 1;
wolffd@0:       end
wolffd@0:     elseif d_class(j) & d_class(j)==c_class(ind2)
wolffd@0:       mj = ind2;
wolffd@0:       mi = ind1;
wolffd@0:       if d_class(j)==c_class(ind1)
wolffd@0:         flag = 1;
wolffd@0:       end
wolffd@0:     end
wolffd@0: 
wolffd@0:     if mj & mi
wolffd@0:       if flag
wolffd@0:         tmp([mj mi],:) = cod([mj mi],:) + epsilon*alpha*...
wolffd@0:                        (dat([j j],:) - cod([mj mi],:));
wolffd@0:       else
wolffd@0:         tmp(mj,:) = cod(mj,:) + alpha * (dat(j,:)-cod(mj,:));
wolffd@0:         tmp(mi,:) = cod(mi,:) - alpha * (dat(j,:)-cod(mj,:));
wolffd@0:       end
wolffd@0:     end  
wolffd@0:   end    
wolffd@0:   inds = find(~isnan(sum(tmp,2)));
wolffd@0:   cod(inds,:) = tmp(inds,:);
wolffd@0: end
wolffd@0: fprintf(1,'\n');
wolffd@0: 
wolffd@0: sTrain = som_set('som_train','algorithm','lvq3',...
wolffd@0: 		 'data_name',data.name,...
wolffd@0: 		 'neigh','',...
wolffd@0: 		 'mask',ones(y,1),...
wolffd@0: 		 'radius_ini',NaN,...
wolffd@0: 		 'radius_fin',NaN,...
wolffd@0: 		 'alpha_ini',alpha,...
wolffd@0: 		 'alpha_type','constant',...
wolffd@0: 		 'trainlen',rlen,...
wolffd@0: 		 'time',datestr(now,0));
wolffd@0: codebook.trainhist(end+1) = sTrain;
wolffd@0: 
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: 
wolffd@0: function nos = class2num(class)
wolffd@0: 
wolffd@0: names = {};
wolffd@0: nos = zeros(length(class),1);
wolffd@0: 
wolffd@0: for i=1:length(class)
wolffd@0:   if ~isempty(class{i}) & ~any(strcmp(class{i},names))
wolffd@0:     names=cat(1,names,class(i));
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: tmp_nos = (1:length(names))';
wolffd@0: 
wolffd@0: for i=1:length(class)
wolffd@0:   if ~isempty(class{i})
wolffd@0:     nos(i,1) = find(strcmp(class{i},names));    
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: 
wolffd@0: 
wolffd@0: