GRNN/PNN vs RBF

Hello Greg,

why isn't there any function like newrb for GRNN/PNN? I mean function which add X neurons every itaration until MSE goal is reached. Maybe I miss something and it's not usefull. So far I split my dataset to different sizes and I test it for different values of spread, do you think it's a good idea?

I think I know how these networks work theoretically. There is one more layer in GRNN, where are two neurons (denominator summation unit and numerator summation unit). My question is, how adding these two neurons improve original RBF network? What is advantage of use GRNN against RBF network?

Answer 1

[Greg Heath]

I do not recall any significant difference between

NEWPNN, NEWGRNN & NEWRBE

They can ALL be replaced by the more flexible NEWRB.

See my NEWSGROUP tutorial on NEWRB

Hope this helps.

Thank you for formally accepting my answer

Greg

PS: searching on greg NEWRBE, greg NEWPNN and greg NEWGRNN may also bring up something useful

[EdWood]

I think there is significiant difference actually. NEWPNN/NEWGRNN are much faster than NEWRBE, they are also less memory expensive. I tried NEWGRNN and NEWRBE on my dataset with 36 inputs and 1600 records.

RBF: Elapsed time is 1.700052 seconds.
GRNN: Elapsed time is 0.048942 seconds.

I repated measure several times. If I am right, there should be definitely some function like NEWRB for GRNN/PNN networks.

Answer 2

[Greg Heath]

NEWRB can be used for both REGRESSION and CLASSIFICATION (I have done both zillions of times).

I cannot think of any case where NEWRB cannot be used instead of FITNET or PATTERNNET. Obviously there are cases where it is not optimal or even much worse. However, that is not the point.

Hope this helps.

Thank you for formally accepting my answer

Greg

[EdWood]

I know NEWRB can be used for regression and classification, but this is not what I am talking about. What I am trying to say is that NEWGRNN is better than NEWRBE, prediction performance is the same i guess but it's less time expensive, which I mentioned above. I think it's also less memory expensive. So I am curious why is there not any function like NEWRB for NEWGRNN, adding new neurons step by step instead of train whole training set at once. Maybe I didn't explain it properly before, sorry for that.