Radial Bases Function Networks with K means
in [Matlab]
|
From: Greg Heath on 28 Jul 2010 21:21 On Jul 28, 7:43 am, "naser " <naser.sepe...(a)gmail.com> wrote: > GregHeath<he...(a)alumni.brown.edu> wrote in message <5e6f9742-fdcb-4e6e-ab05-ee1c3298d...(a)k19g2000yqc.googlegroups.com>... > > On Jul 27, 1:51 pm, "naser " <naser.sepe...(a)gmail.com> wrote: > > > GregHeath<he...(a)alumni.brown.edu> wrote in message <2a928e05-2b64-487f-b32a-e3d4f18e9...(a)k39g2000yqb.googlegroups.com>... > > > > On Jul 25, 8:52 am, "naser " <naser.sepe...(a)gmail.com> wrote: > > > > > HI, > > > > > I want to use a Radial Bases Function Networks with K means (for center determine). > > > > > Would you please tell me how can I do this with matlab function? > > > > > > Best regards > > > > > Do you have theneuralnetwork toolbox? > > > > > Greg > > > > Yes, I have theneuralnetwork toolbox. > > > Would you please help me. > > > If you can be satisfied with identical, spherical cluster shapes, use > > NEWRB, NEWRBE or > > NEWPNN with the training set of cluster centers. However, an > > independent validation set > > is needed to determine the radius of the clusters and an independent > > test set is needed to > > obtain an unbiased estimate of regression or classification error. > > > Hope this helps. > > > Greg > > for examlpe : > p=[1 3;5 10] as input fuction > t=[2 3] as target > > would you please help me how can I use Radial Bases Function Networks with K means for centroid? In general, you need to know more than the just the location of the cluster centers. From the output of the clustering program you would expect the estimation of a priori probabilities, and covariance matrices. For the purposes of discussion, assume that the distributions are equiprobable spherically symmetric Gaussians with the same standard deviation s1=s2=s. If p is the set of cluster centers and t is the corresponding set of outputs. You can use newrb, newrbe or newgrnn for regression or classification. In addition, you can also use newpnn for classification. doc newrb doc newrbe doc newgrnn doc newpnn The input parameter "spread" is problem dependent and, in general, will have to be chosen by trial and error. For classification use t = [1 0] as the training target and assign inputs to class 1 if y = round(sim(net,x)) >= 0.5; otherwise assign them to class 2. Define p1 = p(:,1) and p2 = p(:,2). Then the distance between cluster centers is d12 = sqrt((p2-p1).^2) and the optimal spread in the call of the functions will depend on s but be smaller than d12. Given s you can deduce a good value for spread via trial and error using simulated validation sets P1 = repmat(p1,1,N1)+s1*randn(2,N1); Y1 = sim(net,P1); and similarly for P2. After you have decided on a value for spread, estimate the error using a test set from the same simulation formulas. Good Luck. Greg |