Why the results of my Elman network are different every time?

I created a elman network. But the results every time I run the code were different.I got different "errors","regression" and "avg_error" Could anyone tell me why? Appreciate SO MUCH!

Here is the code.

clear all
load('input4_train.mat');
load('output4_train.mat');
load('input4_test.mat');
load('output4_test.mat');

inputSeries = tonndata(input4_train,false,false);
targetSeries = tonndata(output4_train,false,false);
inputTest = tonndata(input4_test,false,false);
outputTest = tonndata(output4_test,false,false);

% Create a Network
hiddenLayerSize = 5;
net=newelm(inputSeries,targetSeries,[10,3,1], {'tansig','logsig','purelin'});

% Setup Division of Data for Training, Validation, Testing
net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;
net.trainParam.epochs = 2000;

% Initial net
net = init(net);

% Train the Network
net = adapt(net,inputSeries,targetSeries);

% Test the Network
outputs = sim(net,inputTest);

errors = gsubtract(outputTest,outputs);
error = cell2mat(errors);
for i = 1:10
    error(i)=abs(error(i));
end
avg_error = sum(error)/10;
performance = perform(net,outputTest,outputs)
% View the Network
view(net)

% Plots
figure, plotregression(outputTest,outputs)
figure, plotresponse(outputTest,outputs)
figure, ploterrcorr(errors)

[Greg Heath]

1. MATLAB's ELMAN has been obsolete for at least 5 years.

2. What scenario did you use it for? Why was it used instead of timedelaynet or narxnet?

3. Where did you obtain this code? Many of the options chosen make little sense.

4. What is the output of the following whos statement?

 whos inputSeries input4_train targetSeries output4_train inputTest input4_test outputTest output4_test

5. The code has so many inconsistencies and errors, it really isn't worth spending much time on it. Unfortunately the help and doc documentation on ELMAN is very disappointing.

6. If you insist on using it, answer my questions and I will help using MATLAB's simpleseries_dataset as an example.

Hope this helps.

Greg

[Greg Heath]

% load('input4_train.mat');
% load('output4_train.mat');
% load('input4_test.mat');
% load('output4_test.mat');
% 
% inputSeries = tonndata(input4_train,false,false);
% targetSeries = tonndata(output4_train,false,false);
% inputTest = tonndata(input4_test,false,false);
% outputTest = tonndata(output4_test,false,false);

whos

% % Create a Network
% hiddenLayerSize = 5;

Value never used

% net=newelm(inputSeries,targetSeries,[10,3,1], {'tansig','logsig','purelin'});

No justification for 3 hidden layers. One is sufficient.

% % Setup Division of Data for Training, Validation, Testing
% net.divideParam.trainRatio = 70/100;
% net.divideParam.valRatio = 15/100;
% net.divideParam.testRatio = 15/100;

Above 3 commands unnecessary for default values.

% net.trainParam.epochs = 2000;
% 
% % Initial net
% net = init(net);
% 
% % Train the Network
% net = adapt(net,inputSeries,targetSeries);
% 
% % Test the Network
% outputs = sim(net,inputTest);
% 
% errors = gsubtract(outputTest,outputs);
% error = cell2mat(errors);
% for i = 1:10
%     error(i)=abs(error(i));
% end
% avg_error = sum(error)/10;

Above 5 commands unnecessary.

help mae
doc mae

% performance = perform(net,outputTest,outputs)
% % View the Network
% view(net)
% 
% % Plots
% figure, plotregression(outputTest,outputs)
% figure, plotresponse(outputTest,outputs)
% figure, ploterrcorr(errors)

Defaults: Above 3 commands unnecessary;

[Heather Zhang]

Thank you so much for your detailed,Greg! I need your help on my code. I will answer your questions.

Q2. What scenario did you use it for? Why was it used instead of timedelaynet or narxnet? I use it for dealing with a time series prediction issue.

Q3. Where did you obtain this code? Many of the options chosen make little sense. I coded it according to BP network code. I knew "newelm" function is for build an Elman network and "sim" is for testing a network. So I coded it by myself. I can tell there should be something wrong in it, but I don't know where the mistakes are.

So pleas help me on this code, and that would be appreciated very much!

Answer 1

[Walter Roberson]

Neural Networks initialize their weights randomly usually. If you want repeatability you can initialize the weights yourself or you can set the random number generator seed.

[Heather Zhang]

Thank you so much,Walter! I use 'init' function to initialize the weights. Am I right? Looking forward to your answer.

[Greg Heath]

YOU CAN ALSO USE THE FUNCTION CONFIGURE