added very experimental rls code - dont use this for now unless as a …

…template for some other prediciton method!!

rls/rlsPredict.m

@@ -0,0 +1,31 @@
+function Ytest = rlsPredict(Xtest,Xtrain,coeffs,whichKernel)
+%prediction from RLS coeffs from rlsTrain
+% function Ytest = rlsPredict(Xtest,Xtrain,coeffs,whichKernel)
+%
+% returns:
+% Ytest prediction
+%
+% takes
+% Xtest points on which to test
+% Xtrain training dataset
+% coeffs coefficient from rlsTrain
+% whichKErnel kernel to use
+
+rlsAssignkernelfun;
+
+[N,d] = size(Xtrain);
+Nt=size(Xtest,1);
+
+for i=1:Nt
+   % s=0;
+    %for j=1:N
+   %     s=s+coeffs(j)*kernelfun(Xtrain(j,:),Xtest(i,:));
+    %end;
+    Ytest(i)=sum(coeffs.*kernelfun(Xtrain(:,:),Xtest(i,:)));
+    %Ytest(i) = sum((Xtrain,Xtest(i,:)')*coeffs);
+end;
+
+%Ytest = (Xtest*Xtrain')*coeffs;
+
+
+

rls/rlsTrain.m

@@ -0,0 +1,28 @@
+function [coeffs,K] = rlsTrain(Ytrain,Xtrain,whichKernel,lambda,varargin)
+%rlsTrain regularized least squares coefficient estimator
+% function coeffs = rlsTrain(Ytain,Xtrain,whichKErnel,lambda)
+%
+% returns:
+% optimal RLS coefficients in coeffs
+% K Kernel matrix of training set
+%
+% takes
+% Xtrain training input
+% Ytrain training labels
+% whichKErnel kernel to use
+% lambda regularization weight
+% (optional) kernel matrix K
+
+%% compute kernel matrix
+if size (varargin) ==1 % if kernel matrix is supplied
+    K=varargin{1};
+else % compute kernel matrix
+    rlsAssignkernelfun;
+    K = KernelMatrix(Xtrain,kernelfun);
+end;
+
+%% perform RLS fit
+
+[N,d] = size(Xtrain);
+
+coeffs  = ( K+ (eye(N)*lambda) )\Ytrain;