MATLAB Answers

0

Why my network is not giving the desired output

Asked by vaishnavi potharaju on 12 Feb 2019
Latest activity Commented on by Greg Heath
on 21 Feb 2019
I'm trying to design a neural network using nntool of matlab R2015a having input layer of 27 neurons, output layer of 2 neurons and one hidden layer of 10 neurons.I have scaled the input and output data to (0,1) for logsig activation function of hidden layer with purelin in the output layer.For tansig activation function in hidden layer i have scaled the data to (-0.5,0.5). I have trained the network with 1155 training patterns. My mse and R are very good,but the network is not giving the expected result when tested with new data. I have tried almost all the combinations possible from the user inerface(nntool) and in a great confusion. It would be very helpful if answered.
Thank you.

  0 Comments

Sign in to comment.

1 Answer

Answer by Greg Heath
on 18 Feb 2019
 Accepted Answer

  1. Design(training+validation), test and new data should all have the same summary statistics BEFORE NORMALIZATION. This may require mixing all of the data together before creating the train/val/tst subsets.
  2. I prefer zero-mean, unit-variance normalization. It is very helpful for spotting outliers.
  3. One hidden layer of tanh (aka tansig) hidden nodes is sufficient. However, occasionally, problem specifics make using 2 or more to be appropriate.
  4. Use as few hidden nodes as possible to reinforce stability. I start with 1 and use an outer loop to increase the number until the training MSE is less than 0.01 times the training target variance.
  5. An inner loop is used to obtain 10 designs that differ because of random weight initializations.This yields a training RSQUARE of 0.99.
  6. For each sample choice of hidden nodes design 10 or more nets that differ by initial random weights.

  2 Comments

Thank you very much Mr.Greg Heath for your early response. Till now I am using user interface(nntool) of matlabR2015a for my neural network training.
1. From your reply i understood that i have to write code for my network training which im not really good at. Could you please suggest any alternate solution to my problem.
2. I have read that for zero mean and unit variance normalization we should use logsig activation function but according to your reply it would be tansig which is generally used for (-1,1) scaling interval.Which criteria should i follow?
  1. No.
  2. Mine.
Greg

Sign in to comment.