added eval_var based on residual correction

MatlabGP/RRNN.m

@@ -53,7 +53,7 @@
 
                 obj.Wout{i} = lsqminnorm([H{i}, ones(size(H{i},1),1)],R);
 
-                Ri(i) = sum(abs(R));
+                Ri(i,:) = sum(abs(R),1);
 
                 R = R - [H{i}, ones(size(H{i},1),1)]*obj.Wout{i};
 
@@ -103,8 +103,33 @@
             y = obj.unscale_y(y);
         end
 
-        function sig = eval_var(~,x)
-            sig = 0*x(:,1);
+        function sig = eval_var(obj,X)
+            %PREDICT Predicts the output of the trained model for new input
+            %data
+            % Inputs:
+            %   obj - trained RRNN
+            %   X - Input data
+
+            % Output:
+            %   y - output output
+
+            H = obj.activation.forward([X, ones(size(X,1),1)] * obj.Win{1});
+            y1 = [H, ones(size(H,1),1)] * obj.Wout{1};
+
+            y = 0;
+            X = obj.scale(X);
+
+            for i = 1:numel(obj.Win)
+
+                H = obj.activation.forward([X, ones(size(X,1),1)] * obj.Win{i});
+                y = y + [H, ones(size(H,1),1)] * obj.Wout{i};
+
+            end
+
+            y = obj.unscale_y(y);
+            y1 = obj.unscale_y(y1);
+
+            sig = (y1 - y).^2;
         end
 
         function [mu,sig] = eval_all(obj,x)

MatlabGP/examples/testRRNN.m

@@ -3,23 +3,24 @@
 close all
 clc
 
-xx = [0;lhsdesign(10,1);1];
+xx = [0;lhsdesign(8,1);1];
 yy = normrnd(forr(xx,0),0*forr(xx,0));
 
 xmesh = linspace(0,1,100)';
 ymesh = forr(xmesh,0);
 
-nnet = RRNN(NN.SWISH(0.8),10,0,1);
+nnet = RRNN(NN.SWISH(2),10,0,1);
 
 %%
 
 tic
-nnet2 = nnet.train(xx,yy,5);%,xv,fv
+[nnet2,Ri] = nnet.train(xx,yy,5);%,xv,fv
 toc
 
 %%
 
 yp2 = nnet2.eval(xmesh);
+sig2 = nnet2.eval_var(xmesh);
 
 
 %%
@@ -35,6 +36,11 @@
 plot(xmesh,yp2)
 plot(xx,yy,'x')
 
+figure
+plot(xmesh,sig2)
+
+1 - mean((ymesh - yp2).^2)./var(ymesh)
+
 %%
 
 function y = forr(x,dx)