Organization: Information Technology Institute, National Computer Board, S'pore
Date: Mon, 7 Sep 1992 01:53:09 GMT
Lines: 32
Hello fellow netters,
I am trying to use RBF-based hybrid network for prediction. I follow the
design as is given in literature - Single hidden layer feed forward BP-net
with linear output nodes and guassian hidden nodes. I also use additional nodes each of input dimension for every hidden unit (introduced for improving prediction). I have simulated this net using around 100 guassians for a 12 dimensional input of around 400 input sample size (could be oversized - I don't know).
I have some fundamental questions regarding this :
1) The unsupervised part of learning in this context is basically a process of clustering all the input patterns of N-dimension. The common way of quantifying the tightness of clustering is sum((Eu)dist of each pattern from its closest cluster center). What should this measure be for a good clustering? OR in other words "Is there a bench mark against which the quality for a given clustered sample could be compared ( such as Average Relative Variance for prediction)"
2) Is there any empirical relation between no: of attractors (cluster centres), input pattern dimension and no: of patterns to be clustered ?
3) Any figures of reasonable input dimension for which RBF network size will be manageble ?
I request for answers to these queries . Also am thankful for any pointers to literature discussing these issues. I shall summarise the replies I get on this news-group.
PS: To my previous posting of this note, to my dissapointment I received no replies. Has the interest on RBF nets died out? If so, could any one summarise
