RBF newrbe algorithm uses k-means and inverse matrix?

Does newrbe algorithm use k-means or random vectors for neuron center? For weights calculations is uses inverse matrix?

Answer 1

[Greg Heath]

No. NEWRBE constructs identical symmetric Gaussians around EACH data point. Therefore, to optimize the design, just vary the common radius of the Gaussians.

Weights are determined using pseudo-inversion via the minimum-mse BACKSLASH solution x = A\b to the linear matrix equation A*x = b.

For serious RBF work, consider a modification of NEWRB where Gaussians are iteratively constructed from the poorest performing data point.

Years ago I made some substantial improvements to the NEWRB algorithm in the NEWSGROUP. Try searching backwards in time with the search words

 greg newrb

Hope this helps.

Thank you for formally accepting my answer

Greg

[Xiaoran Li]

Hi, Greg. I found you answered many questions on RBFNN. And I want to ask a question similar like this. 'I am studying on RBFNN, and I have read about K-mean, OLS to determine the center. But when I use the function newrb, there this no need to determine how to choose the center. I want to know how does the function newrb choose data center, and how it works.' I will be greatly appreciated for your help! Thanks!

[Greg Heath]

I just answered that in the previous post:

The input with the highest error is added to the training set.