nn2train.m

神经网络设计一些资料

function tweights=NN2train(IWeights,TDIN,TDOUT,LR,ErTh,AF)
[rows,numdatasets]=size(TDIN);
toterr=ErTh+1;
epoch=0;
while(toterr>ErTh)
    disp(toterr);
    dataorder=randperm(numdatasets);
    for n=1:numdatasets
        output=NN2(IWeights,TDIN(:,dataorder(n)),AF);
        %Should update bias weights
        for j=1:IWeights.outputnum
            if(AF=='p')
                IWeights.bias(j)=IWeights.bias(j)+LR*(TDOUT(j,dataorder(n))-output(j));
            else
                IWeights.bias(j)=IWeights.bias(j)+LR*output(j)*(1-output(j))*(TDOUT(j,dataorder(n))-output(j));
            end
        end
        for r=1:IWeights.outputnum
            for s=1:IWeights.inputnum
                if (AF=='p')
                    IWeights.inputs(r,s)=IWeights.inputs(r,s)+LR*(TDOUT(r,dataorder(n))-output(r))*TDIN(s,dataorder(n));
                else
                     IWeights.inputs(r,s)=IWeights.inputs(r,s)+LR*(TDOUT(r,dataorder(n))-output(r))*TDIN(s,dataorder(n))*output(r)*(1-output(r));
                 end
             end
         end
     end
     %should be finished for an epoch
        temperr=0;
        for n=1:numdatasets
            output=NN2(IWeights,TDIN(:,dataorder(n)),AF);
            %disp(output)
            %display(TDOUT(:,dataorder(n))')
            %disp(TDOUT(:,dataorder(n))-output')
            temperr=temperr+sum((TDOUT(:,dataorder(n))-output').^2);
        end
        toterr=temperr/(IWeights.outputnum*2);
    epoch=epoch+1;    
    end
    disp(epoch);
    tweights=IWeights;