nnmodel.asv

人工神经网络：MATLAB源程序用于训练测试

					else
						nois = noilev*randn(size(xtr));
					end
					xtrnois = xtr+nois;
	
					for i = 1:mxtr
						xtrunc = 0*xtrnois;		% All training input variables but one
						xtrunc(i,:) = xtrnois(i,:);	% are set to 0

						% Training of the network. The monitoring set is used to stop the training
						[w1n,w2n] = levmarq(topo,w1inref,w2inref,xtrnois,ytr,xm,ym,epochs,[]);

						% Estimation of training responses using only one input in the model
						[ytrunc1,ytrunc2] = lmeval(topo,w1n,w2n,xtrunc);
	
						% Inverse-scaling of training responses estimated using one input only
						ytrunchat = invrange(ytrunc2,0.2,0.8,tableout);
						concyhtr2(i,:) = ytrunchat;					
					end
					% The replicate y-values predicted with partial models are accumulated  
					conctrx = [conctrx;xtr'];		% Matrix of replicate training x-values
					concytrhat = [concytrhat;concyhtr2'];	% Matrix of replicate training y-values
				end

				% For each input variable, an ANOVA-LOF is performed on the replicate y-values to
				% check if they fit a linear regression model
				for i = 1:mxtr
					[ylhat,F,Ftab,lof,b0,b1] = linfit(conctrx(:,i),concytrhat(:,i));
					FF(i) = F;		% The calculated F-value for each variable
					loff(i) = lof;		% The ratio F/Fcritical for each variable
				end
				% Kolmogorov-Smirnov test to check that for each sample, replicate 
				% predicted values are normally distributed.
				resultks = kolnorm(concytrhat,nxtr,32,0.05);
				if any(resultks)
					disp(' ANOVA for lack-of-fit may not be reliable')
					disp(' Some replicate predicted values are not normally distributed')
					disp(' ')
					disp(' Percentage of samples not normally distributed for each variable:')
					disp(' -----------------------------------------------------------------')
					disp(resultks)	
					men3 = menu('ANOVA-LOF not reliable',...
					'Continue');
				end	
				bar(FF);
				nax = axis;hold;clc;
				line([nax(1) nax(2)],[Ftab Ftab],'linestyle','--','color','r');
				xlabel('Variable');ylabel('ANOVA F-value');
				set(gcf,'Position',setfig);
				men3 = menu('ANOVA for lack-of-fit',...
				'1. Print',...
				'2. Next plot');
				if men3 == 1
					print
				end;clc;close;close
			elseif men2 == 5
				continue = 0;
			end
		end
	end

	if men1 == 5	% Partial models (hidden nodes)
		clc;close
		for i = 1:mdef
			rhn = corrplot(ytrain,partyhh(i,:)',1);	% Correlation plot
			xlabel('Observed values - Training set')
			rensy = sprintf('Predicted values when hidden node %2.0f is removed',i);
			ylabel(rensy)
			set(gcf,'Position',setfig);
			men2 = menu('Partial model - Training',...
			'1. Print',...
			'2. Next plot');
			if men2 == 1
				print
			end;close
		end
	end

	if men1 == 6	% Projection on hidden nodes
		clc;close
		continue = 1;
		while continue
			clc;close
			men2 = menu('Projection on hidden nodes',...
			'1. Training set only',...
			'2. Training + monitoring set',...
			'3. Training + monitoring + test set');
			men3 = menu('Display',...
			'1. Points only',...
			'2. Show sample indices');
			node1 = input('Which node on X-axis ?    ');
			node2 = input('Which node on Y-axis ?    ');
			plot(ytr1(node1,:),ytr1(node2,:),'o')
			hold;clc
			rensx = sprintf('Outputs of hidden node %2.0f (%3.0f%%)',[node1 varytrunch(node1)]);
			rensy = sprintf('Outputs of hidden node %2.0f (%3.0f%%)',[node2 varytrunch(node2)]);
			xlabel(rensx);ylabel(rensy);
			title('o: Training set')
			if men2 ~= 1
				plot(ymon1(node1,:),ymon1(node2,:),'c*');
				title('o: Training set   * : Monitoring set')
				if men2 == 3
					if testflag
						plot(ytes1(node1,:),ytes1(node2,:),'r+');
						title('o: Training set   * : Monitoring set   +: Test set');
					else
						title('o: Training set   * : Monitoring set   No test set');
					end
				end
			end
			if men3 == 2
				for i = 1:nxtr
					intr = int2str(i);
					textr = ['t' intr];
					text(ytr1(node1,i),ytr1(node2,i),textr);
				end				
				if men2 ~= 1
					for i = 1:nxm
						inm = int2str(i);
						texm = ['m' inm];
						text(ymon1(node1,i),ymon1(node2,i),texm);
					end
					if men2 == 3
						if testflag
							for i = 1:nxte
								inte = int2str(i);
								texte = ['n' inte];
								text(ytes1(node1,i),ytes1(node2,i),texte);
							end
						end
					end				
				end
			end
			set(gcf,'Position',setfig);	
			men4 = menu('Projection on hidden nodes',...
			'1. Print',...
			'2. Another plot',...
			'3. Back to main menu');
			if men4 == 1
				print
			elseif men4 ==3
				continue = 0;
			end;close
		end
	end

	if men1 == 7	% Calculation of deviations from linearity
		clc;close
		continue = 1;
		while continue
			men2 = menu('Deviations from linearity',...
			'1. Training set',...
			'2. Monitoring set',...
			'3. Test set',...
			'4. Back to main menu');
			 if men2 == 1		% Training results
			 	for i = 1:mdef
					maxintr = 2*max(abs(intr1(i,:)));
					axinp = [-maxintr:0.1:maxintr];
					projmaxinp = pmntanh(axinp);
					clc;close
					devlin(i,:) = (ytr1(i,:)-intr1(i,:))./intr1(i,:);
					meandev = round(100*mean(abs(devlin(i,:))));
					plot(axinp,axinp,'r',axinp,projmaxinp)
					hold;clc;
					plot(intr1(i,:),ytr1(i,:),'co')
					renstit = sprintf('Hidden node %2.0f - magnitude %3.0f%% - Mean deviation: %3.0f%%',[i varytrunch(i) meandev]);
					title(renstit)
					ylabel('Output')
					xlabel('Input - Training set')
					set(gcf,'Position',setfig);
					men3 = menu('Deviations from linearity',...
					'1. Print',...
					'2. Next plot');
					if men3 == 1
						print
					end;close			
				end
				clear devlin
			elseif men2 == 2	% Monitoring results
				for i = 1:mdef
					maxintr = 2*max(abs(inmon1(i,:)));
					axinp = [-maxintr:0.1:maxintr];
					projmaxinp = pmntanh(axinp);
					clc;close
					devlin(i,:) = (ymon1(i,:)-inmon1(i,:))./inmon1(i,:);
					meandev = round(100*mean(abs(devlin(i,:))));
					plot(axinp,axinp,'r',axinp,projmaxinp)
					hold;clc;
					plot(inmon1(i,:),ymon1(i,:),'co')
					renstit = sprintf('Hidden node %2.0f - magnitude %3.0f%% - Mean deviation: %3.0f%%',[i varytrunch(i) meandev]);
					title(renstit)
					ylabel('Output')
					xlabel('Input - Monitoring set')
					set(gcf,'Position',setfig);
					men3 = menu('Deviations from linearity',...
					'1. Print',...
					'2. Next plot');
					if men3 == 1
						print
					end;close			
				end
				clear devlin
			elseif men2 == 3	% Test results
				if testflag
					for i = 1:mdef
						maxintr = 2*max(abs(intes1(i,:)));
						axinp = [-maxintr:0.1:maxintr];
						projmaxinp = pmntanh(axinp);
						clc;close
						devlin(i,:) = (ytes1(i,:)-intes1(i,:))./intes1(i,:);
						meandev = round(100*mean(abs(devlin(i,:))));
						plot(axinp,axinp,'r',axinp,projmaxinp)
						hold;clc;
						plot(intes1(i,:),ytes1(i,:),'co')
						renstit = sprintf('Hidden node %2.0f - magnitude %3.0f%% - Mean deviation: %3.0f%%',[i varytrunch(i) meandev]);
						title(renstit)
						ylabel('Output')
						xlabel('Input - Test set')
						set(gcf,'Position',setfig);
						men3 = menu('Deviations from linearity',...
						'1. Print',...
						'2. Next plot');
						if men3 == 1
							print
						end;close			
					end
				else
					men3 = menu('No test set was provided',...
					'Back to previous menu');
				end
				clear devlin
			elseif men2 == 4
				continue = 0;
			end
		end
	end

	if men1 == 8	% Summary results
		clc;close
		disp(' Optimal number of epochs for each trial:')
		disp(' ----------------------------------------')
		disp(indisp)
		disp(' ')
		disp(' ')				
		disp(' Seeds:')  
		disp(' ------')
		disp(rseed')
		disp(' ')
		disp(' ')
		disp(' Seed for the solution retained:')  
		disp(' -------------------------------')
		disp(rseed(indref))
		disp(' ')
		disp(' ')
		disp(' RMSEC  -  RMSEM:')
		disp(' ----------------')
		disp([rmstra' rmsmon']) 
		disp(' ')
		disp(' ')
			
		if trials ~= 1
			disp(' Median of RMSEM:')
			disp(' ----------------')
			disp(medrmsmon)
			disp(' ')
			disp(' ')
			disp(' Average RSMSEC  -  Average RMSEM:')
			disp(' ---------------------------------')
			disp([mean(rmstra) mean(rmsmon)])
			disp(' ')
			disp(' ')
			disp(' STD(RMSEC)  -  STD(RMSEM):')
			disp(' --------------------------')
			disp([std(rmstra) std(rmsmon)])
			disp(' ')
			disp(' ')
			if testflag
				disp(' RMSEP')
		 		disp(' -----')
		 		disp(rmstest')
				disp(' ')
				disp(' ')
				disp(' Average RMSEP:')
				disp(' --------------')
				disp(mean(rmstest))
				disp(' ')
				disp(' ')
				disp(' STD(RMSEP):')
				disp(' -----------')
				disp(std(rmstest))
				disp(' ')
				disp(' ')
			end
		end
		[sensort,order] = sort(varytrunc);	% Sensitivities ranked in ascending order
		disp(' Variables ranked in order of increasing sensitivity:')
		disp(' ----------------------------------------------------')
		disp(order)
	end

	if men1 == 9	% Draw the neural network
		clc;close
		drawnn(topo,w1f,w2f);
		set(gcf,'Position',setfig);
		title('Topology of the neural network');
		xlabel('Nodes with vertical bars indicate linear transfer functions')
		men2 = menu('NN structure',...
		'1. Print',...
		'2. Back to main menu');
	 	if men2 == 1
	 		print
		end;close
	end

	if men1 == 10	% End
		contflag = 0;
		clc;close
	end
end
w1f = w1ref;
w2f = w2ref;
clear concymhat concyteshat partyhh conctrx concytrhat concyhtr concyhtr2