"aarti gehani" wrote in message <khl33o$rbp$1@newscl01ah.mathworks.com>...
> I am not clear with the right selection of spread constant in RBFN.
>
> When I selected spread constant = 0.1, the performance goal was reached but the error was large but when I selected spread constant = 10, the performance goal was not reached and error was very small.
>
> Can anyone tell me that if I consider spread constant = 10, then the results will be correct?
Glossing over minute details:
1. Standardize your inputs and targets using zscore (zeromean/unitvariance)
2. Randomly divide your data into a training set, a hold out validation set and a test set.
3. Set a training goal of MSEtrngoal = 0.01*mean(var(ttrn'))
4. Loop over 10 or more spread value candidates.
5. Choose the best net using MSEval
6. If unsatisfactory, go back to 4, otherwise
5. Obtain a final evaluation on the validation set choice using the test set.
7. If unsatisfactory, go back to 2
For more gory details than you anticipated, go to
http://www.mathworks.com/matlabcentral/newsreader/view_thread/151286#888395
Hope this helps
Greg![]()
> I am not clear with the right selection of spread constant in RBFN.
>
> When I selected spread constant = 0.1, the performance goal was reached but the error was large but when I selected spread constant = 10, the performance goal was not reached and error was very small.
>
> Can anyone tell me that if I consider spread constant = 10, then the results will be correct?
Glossing over minute details:
1. Standardize your inputs and targets using zscore (zeromean/unitvariance)
2. Randomly divide your data into a training set, a hold out validation set and a test set.
3. Set a training goal of MSEtrngoal = 0.01*mean(var(ttrn'))
4. Loop over 10 or more spread value candidates.
5. Choose the best net using MSEval
6. If unsatisfactory, go back to 4, otherwise
5. Obtain a final evaluation on the validation set choice using the test set.
7. If unsatisfactory, go back to 2
For more gory details than you anticipated, go to
http://www.mathworks.com/matlabcentral/newsreader/view_thread/151286#888395
Hope this helps
Greg