k_mean.m

The neuro-fuzzy software for identification and data analysis has been implemented in the MATLAB lan

function [centers,bases]=k_mean(comp,training_set)

%K_MEAN
%  [centers,bases]=k_mean(comp,training_set)
%
%	centers[no_rules,no_var]   centers of the membership functions
%	bases[no_rules,no_var]	   bases   of the membership functions   	
%
%	comp			complexity (number of fuzzy rules)
%	training_set		the training set matrix
%
	
%___________________________________________________10 April 1996_______


global memb_fun;

eps=0.0;

no_var=size(training_set,2)-1;
centers=zeros(comp,no_var);
ranges=zeros(no_var,2);

for j=1:no_var
	ranges(j,1)=min(training_set(:,j))-eps;
	ranges(j,2)=max(training_set(:,j))+eps;
	
	centers(:,j)=ranges(j,1)+...
		rand(comp,1)*(ranges(j,2)-ranges(j,1));
end


data=training_set(:,1:no_var);			
[no_data,no_variables]=size(data);

change=1;
data_norm=zeros(size(data));
for j=1:no_var
	data_norm(:,j)=Normalized(data(:,j),ranges(j,1),ranges(j,2));
end


while (change >0.001)
	[no_clusters,no_var]=size(centers);
	no_data_cluster=zeros(no_clusters,1);
	change=0.0;
	no_new_centers=0;
	centers_norm=zeros(size(centers));
	stdev=zeros(size(centers));
	bas_max=zeros(size(centers));
	minimum=zeros(size(centers));
	maximum=zeros(size(centers));
	

	for j=1:no_var
		centers_norm(:,j)=Normalized(centers(:,j),ranges(j,1),ranges(j,2));
	end

	distances=dist(data_norm,centers_norm'); 
					% distances [ no_data,no_clusters]
	if no_clusters>1
		[a,cluster_related]=min (distances');
	else
		cluster_related=ones(no_data,1);
	end
	
	for i=1:no_clusters
		no_data_cluster(i)=sum ( cluster_related == i);
	end
	data_clustered=zeros(size(data));

	for i=1:no_clusters
		if no_data_cluster(i)>0
			no_new_centers=no_new_centers+1;
			start_data=sum(no_data_cluster(1:i-1))+1;
			end_data=sum(no_data_cluster(1:i));
			data_clustered(start_data:end_data,:)=data(cluster_related == i,:);
			if (no_data_cluster(i)>1)
				new_center=mean(data_clustered(start_data:end_data,:));
				minimum(no_new_centers,:)=min(data_clustered(start_data:end_data,:));
				maximum(no_new_centers,:)=max(data_clustered(start_data:end_data,:));
				stdev(no_new_centers,:)=std(data_clustered(start_data:end_data,:));
				bas_max(no_new_centers,:)=max(abs(new_center-minimum(no_new_centers,:)),abs(maximum(no_new_centers,:)-new_center));
				
			else
				new_center=data_clustered(start_data:end_data,:);	
				stdev(no_new_centers,:)=(ranges(:,2)-ranges(:,1))'/20;
				bas_max(no_new_centers,:)=(ranges(:,2)-ranges(:,1))'/10;
			end
			centers(no_new_centers,:)=new_center;
			change=change+dist(new_center,centers(i,:)');
		end
	end
	centers=centers(1:no_new_centers,:);
	stdev=stdev(1:no_new_centers,:);
	bas_max=bas_max(1:no_new_centers,:);
	no_data_cluster=no_data_cluster(1:no_new_centers);
end


if strcmp(memb_fun,'gaussian')
	bases=stdev.^2;
end

if strcmp(memb_fun,'triangular')
	bases=2*bas_max;
end