The simulate function pre-process the input before computing the prediction. It uses the training set parameters to pre-process a test sample or set.  Check the help of the mapminmax function....

Hope this help.

Fernando

http://www.mathworks.com/help/nnet/ref/mapminmax.html

The simulate function pre-process the input before computing the prediction. It uses the training set parameters to pre-process a test sample or set.  Check the help of the mapminmax function....

Hope this help.

Fernando

http://www.mathworks.com/help/nnet/ref/mapminmax.html

% Choose Input and Output Pre/Post-Processing Functions For a list of all processing functions type: help nnprocess.

The function mapminmax scales inputs and targets so that they fall in the range[–1,1].

the code for normalize data in the range [-1,1]:

[pn,ps] = mapminmax(X);

[tn,ts] = mapminmax(Y);

hi

best is that you regression method in the neural network.

also inorder to improve the results , you can use back-propagation method.

Even more you want accurate results,then use optimization of techniques method.

Hope this helps you a lot.

BEST,

Dr. Indrajit Mandal

Hi,

Well it happened because when we simulate using tool box the weights are not in our control. The system takes default small values and weights are updated during the training process. To control the weights use the random integer generator function 'rng' in the generated code. That way you can preserve the network for future use. I agree with other answers that backpropagation improves the results. If the problem still persists then go for regularization methods such as 'Bayesian Regularization' or try changing the training algorithm.

i hope this helps 

Thanks and regards,

Atiya

dear Majid Nazeer,

In fact I have the same problem....

using mampminmax does not help....

did you get the answer?

can you please help?

Dear Selvaraj,

Thanks for your message, actually I have tried a lot but could not speculate how can I know the equation structure, which a Neural Network uses for estimating a certain value. 

Please do share if you find any resources in this regards.

Thanks

Dear Majid and Raja

I have the same problem, did you solve yours problems. Please help me in this matter.

Dear all,

I have the same problem. Actually, I achieved some very close results, but still not the same as the sim(.). For example, the average difference between the Matlab's output and the by-hand calculation is about 0.017. The major problem is the error per sample in by-hand methodology is bigger than the Matlab's one.

The question about the use of mapminmax is really relevant. In fact, my procedure is:

[ninput,PS] = mapminmax(input); %normalize [-1,1] and save the normalization parameters

%input layer

h11 = dotprod(Wt1,ninput); %dot product between the weights and normalized input

h11 = h11+(b1*ones(1,N_samples)); %add bias

%%%%ook, ok, you can do it in one sentence, but I have done something like this ;)

th11 = tansig(h11);

%output layer

h22 = dotprod(Wt2,th11);

h22 = h22+(b2*ones(1,N_samples));

th22 = purelin(h22); %optional

x1_again = mapminmax('reverse',th22,PS); %use the parameters of the first normalization in order to de-normalize the output data

If someone finds the error...

Well, the problem is solved.

It is all about the normalization...

[pn,ps] = mapminmax(train(:,trInd)); %normalize and get the parameters of the train data

[tn,ts] = mapminmax(target(:,trInd));%normalize and get the parameters of the target values of train data

ninput = mapminmax('apply',input,ps); %apply the normalization of test input according to the parameter ps

%input layer

h11 = dotprod(Wt1,ninput); %dot product between the weights and normalized input

h11 = h11+(b1*ones(1,N_samples)); %add bias

%%%%ook, ok, you can do it in one sentence, but I have done something like this ;)

th11 = tansig(h11);

%output layer

h22 = dotprod(Wt2,th11);

h22 = h22+(b2*ones(1,N_samples));

th22 = purelin(h22); %optional

x1_again = mapminmax('reverse',th22,ts);

----

Look with attention to http://www.mathworks.com/help/nnet/ref/mapminmax.html

 Hi Ana,

Thank you for the explanation. 

If the output layer has one neuron, then all the elements in that output column will be -1. would this affect the training and the accuracy of the ANN? Or should the target be in the same matrix with training matrix as last column and then I separate them before training. 

Thank you,

Aziz

Hi Majid.....

I hope you read all the above comments for the solution but you have still problem. 

When I started my work, I had the same problem...after finding the pre-processing direction.....and how to use it in ANN Matlab code,that's another problem,

Now I completely modified my code with the provided guidelines. In my code I bypass the pre-processing and assigned the initial  weights and bias values =1. Please normalize all your input data between 0.1 and 0.9.

please use the following code for the ....then calculate your answer with ANN prediction hope you will get the same.

Input=in';

Trg=t';

N=5;

T=1;

H1=8;

H2=8;

net = newff(Input,Trg,{H1,H2},{'logsig','tansig','tansig'},'trainlm','learngdm','mse',{},{},'dividerand');

net.trainParam.epochs= 100; %Max no of epochs to train default=1000,

net.trainParam.goal= 0; %Performance goal default=0,

net.trainParam.max_fail= 6; %Max validation fail default=25,

net.trainParam.min_grad= 1e-10; %Max performance gradient default=1e-7;

net.trainParam.mu= 0.001; %Initail Mu default=0.001,

net.trainParam.mu_dec= 0.1; %mu decrease factor default=0.1,

net.trainParam.mu_inc= 10; %mu increase factor default=10,

net.trainParam.mu_max= 1e10; %Max mu default=1e10;

net.trainParam.show= 25; %Epochs b/w displays default=25,

net.trainParam.time= inf; %default

net.divideParam.trainRatio= 0.60;

net.divideParam.valRatio= 0.20;

net.divideParam.testRatio= 0.20;

nn.inputs{1}.processFcns={};

nn.outputs{1}.processFcns={};

net.iw{1,1}=ones(H1,N);

net.lw{2,1}=ones(H1,H2);

net.lw{3,2}=ones(T,H2);

net.b{1,1}=ones(H1,T);

net.b{2,1}=ones(H2,T);

net.b{3,1}=ones(T,T);

net= init(net);

[net,tr]=train(net,Input,Trg);

Result=net(Input)';

Regards 

Hi Afaq, many thanks for the suggestion.

Hi Afaq Ahmad,

please I have few questions a bout what you have posted which looks interesting to me.

1- is N the number of input variable (neurons)?

2- is it necessary to initiate the network?  [net= init(net);]

3- what Result=net(Input)'; does?

4- what the ones matrices needed for?

Thank you

Hi Abdulaziz

Please find the following the answers;

1=Yes N=number of input neurons

2=Yes it is necessary if you delete this from the code or allows only 1 epochs then no weights/bias will be updated.

3=Results are the NN output ...I wanted to save under this name....you can use anything

4=In this code what I assigned the initial weights and bias as ONES. while if you delete this thing from the code then matlab assign the weights and bias randomly.

Please feel free for any feedback/questions.

Regards 

Ahmad

Thank you Afaq fro your concern and  prompt respond

I think you can add the following command before you prompt to go for net command line preparation to make sure that every time you are assigning weights = 1. In this case, no by pass is necessary from per-processing, I guess.

%% Many of the techniques that we are using make use of random sampling. Setting that randstream will allow us to reproduce these results precisely.

s = RandStream('mt19937ar', 'seed', 1);

RandStream.setGlobalStream(s);

Any update/more explanation in this regard would be highly appreciated from the experts.