feedforward net for regression.

I have 1000 sets of data. each set consists of 4 input variables and one output variable. Variables 1, 2, and 3 remain same but only variable 4 changes for one set of data that has 10 records. So, total data size is 10000 x 5. As mentioned variables 1,2 and 3 only change from set to set and within a set they remain constant or same.

How would I use this data to train.

Here is an e.g.

.2, .4, .5, .8 ..... 12

.2, .4, .5, .2 ------ 7

-------------- 10 lines

.1, .6, .8, .5 -------- 10

.1, .6, .8, .8 ---------- 6.3

---------------------- 10 lines

and so on..

total 10000, by 5 Matrix.

Thanks in advance.

[Abolfazl Chaman Motlagh]

You want to use a RNN (recurcive nn) for 10 line, or each 1x4 is an input?

[Pappu Murthy]

I want to use just the feedforward simple net. with input layer and a couple of hiddenlayers and an output layer in general

[Abolfazl Chaman Motlagh]

I recommend you to split data :

X = data(1:4,:);
Y = data(5,:);

and just type :

nftool

the mathwork work on it perfectly i think. and it has all documentation in it. i hope it is what you are looking for.

[Pappu Murthy]

the problem still remains. for each set only one of the input variables (var. 4) is changing and the remaining three are same. So how would you address that?

Answer 1

[Swetha Polemoni]

Hi Pappu Murthy,

It is my understanding you wanted to train a network given input and true lables(Supervised learning). I understand that you have 1000 sets of data. Each set has 4 input variables ,1 output variable and in total you have records in each set. Though the first 3 variables doesn't change within the set, these variables are also important to the network as much as the 4th variables.

This is the example data set you have shared

2, .4, .5, .8 ..... 12

.2, .4, .5, .2 ------ 7

-------------- 10 lines

.1, .6, .8, .5 -------- 10

.1, .6, .8, .8 ---------- 6.3

---------------------- 10 lines

Suppose you ignore the first 3 values while training, the output of the network can be incorrect. For example if the input for the network is as follows

.1, .6, .8, .2

The output of the network can be 7 since we have ignored the first 3 values. Here the network considers only the fourth value i.e., .2 and gives the output 7. So to avoid such errors, all the input variables should be equal importance.

The following documentation might help you

Create and Train a Feedforward Neural Network

[Pappu Murthy]

I understand what you are saying. So this means eventhough the first three inputs are same and only 4th one is varying, can I include them all and train the net? Or would that pose some mathematical problems later because the matrix columns with in a set are same. I really do not know whether it is ok to do so. Is there a way to partition them or something along those lines? I will read the link you provided mean while. Thanks.