Neural Network for Classification: Different Results - same parameters

Question

Benedikt Wessel 2020-6-7

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/544010-neural-network-for-classification-different-results-same-parameters

Bildschirmfoto 2020-06-07 um 21.24.26.png

Hello,

attached you can find a few plots representing the results from my neural network. What is the main reason for the mistake on the rigth site? Is it normal that the lines are seperating different in plot 1 and 2? Maybe the reason is the number of iterations, the learning factor or a bad backpropagation? This is my code:

Any improvements regarding the code are highly appreciated.

clear
clc
%% Input
x1=0.1.*rand(10,2)+0.6;
x2=0.1.*rand(10,2);
x31=0.1.*rand(10,1);
x32=0.1.*rand(10,1)+0.6;
x3=[x31,x32];
x41=0.1.*rand(10,1)+0.6;
x42=0.1.*rand(10,1);
x4=[x41,x42];
x0=-ones(40,1); % negative bias for every Input
in=[x1;x2;x3;x4];
in=[x0,in]'; % complete Inputmatrix
clear x1
clear x2
clear x0
o1=ones(10,1);
o2=zeros(10,1);
output=[o1;o1;o2;o2]'; % correct Output
clear o2
clear o1
weights=2*rand(3,3)-1; % randomn values for the weight matrix
iterations=1000;
coeff=0.5;
err=zeros(iterations,1);
%% Training
for i=1:iterations
    
    out=zeros(40,1);
    numIn=length(in);
    e=0;
    for j=1:numIn
        
        input=in(:,j);
        desired_out=output(:,j);
        
        % Forward-Propagation
        % Hidden_Layer
        H1=dot(input,weights(:,1)); % Hidden-Layer_Neuron1
        H2=dot(input,weights(:,2)); % Hidden-Layer_Neuron2
        HL(1)=sigmoid(H1); % Acitvation function (first Input for the Output_Layer)
        HL(2)=sigmoid(H2); % Activation function (second Input for the Output_Layer)
        % Output-Layer
        input_OL=[input(1);HL(1);HL(2)]; % Input vector for the Ouput-Layer (Bias and the output from the Hidden-Layer)
        O3=dot(input_OL,weights(:,3)); % Output-Layer
        OL=sigmoid(O3); % Output value  neuronal network
        out(j)=OL;
        
        
        % Backward-Propagation
        % Output-Layer
        Hout=[-1,HL]'; % Output  Hiddenlayer
        delta=desired_out-OL; % Deviation  Errorfunction MSE 0.5*(delta)^2
        dsigmoid_OL=OL*(1-OL); % Deviation  Sigmoid-Function (Activationfunction)
        % dO3/dw=H0, Deviation weighted sum regarding the weights 
        weights(:,3)=weights(:,3)+coeff*delta*dsigmoid_OL.*Hout;
        % Hidden-Layer
        dsigmoid_HL=Hout(2:3).*(1-Hout(2:3));
        
        % weigths(input->hidden)+coeff*der(err.fct.)*der(act.fct_Outputlayer)*Input Outputlayer*weigths(hidden->outputlayer)*der(act.fct._Hiddenlayer)*Input
        weights(:,1)=weights(:,1)+coeff*delta*dsigmoid_OL*HL(1)*weights(2,3)*dsigmoid_HL(1).*input; % Weights  1. Neuron in Hidden-Layer
        weights(:,2)=weights(:,2)+coeff*delta*dsigmoid_OL*HL(2)*weights(3,3)*dsigmoid_HL(2).*input; % Weights  2. Neuron in Hidden-Layer
        
    end
    e=e+abs(delta);
    err(i)=e;
end
weights
out
e
%% Plot
plotpv(in(2:3,:),output)
hold on
x=(0:0.1:1);
a=weights(2,1)/weights(3,1);
b=weights(1,1)/weights(3,1);
y=-a*x+b; % aus den Gewichten eine Funktion ableiten
plot(x,y) % diese Funktion grafisch darstellen
a=weights(2,2)/weights(3,2);
b=weights(1,2)/weights(3,2);
y=-a*x+b; % aus den Gewichten eine Funktion ableiten
plot(x,y) % diese Funktion grafisch darstellen
%% Sigmoid
function s=sigmoid(x)
s=1/(1+exp(-x)); 
end