Date: Dec 6, 2012 7:12 PM
Author: Greg Heath
Subject: Re: RBFNN
"dean86" wrote in message <firstname.lastname@example.org>...
> "Greg Heath" <email@example.com> wrote in message <firstname.lastname@example.org>...
> > "dean86" wrote in message <email@example.com>...
> > > Hi all,
> > > I have the following problem:
> > > 777 input from some different sensors, I have to do a PCA and then a RBF to predict the number of analytes.
> > What type of sensors? Unfortunately the trace that I have doesn't specify the type of sensors, it says just "Different kind of sensors".
> > How many measurement vectors for NN input? N = 343? Yes
> > How many dimensions in each input vector? I = 777 ? Yes
> > Regression or classification ? I'm sorry but actually I just know that I have to use the scores of the PCA to train a radial basis function neural network to predict concentration of the analytes. I'm sorry but once again the web is the only resource that I have to learn this topic.
> > What is the output ?
> > How many dimensions in each output vector? O = ?
> > > Could you help me to understand the correct way to do the RBF, to use the command 'newrb', but above all, how to understand the results?
> > > First thing that I do is to load the data, then I have this matrix 777x343, so I do the transpose and I start to do the mean-centring and then the PCA on this matrix and I obtain the scores (343x4) and the loadings (777x4). Now I have to use this scores to do this RBF, so I obtain the transpose of the scores matrix (4x343) and now, should I use the newrb with this last matrix and the original data matrix (777x343)?
> > >
> > What is the criterion for input dimensionality reduction why 777==> 4? This makes no sense to me. Because when I use the PCA I can clearly see from the plot that I can keep just 4 variables.
> > All NEWRB needs are input and output matrices and a reasonable value for the MSE goal and a range of candidate spread values.
> > For regression standarize both input and target matrices. For c-class classification, standardize the input matrix but use one of c (=O) binary coding for the output matrix.
> > Use MSEgoal = 0.01*mean(var(t',1)) % yields R^2 >= 0.99
> > Obtain multiple designs from a loop over spread values. I usually start with a coarse search spread = 2^(i-1), i = 1,2,... Then refine the search if needed.
> > Old posts:
> > 5 threads for heath newrb overfitting overtraining
> > Neural Networks Question
> > Newrb with k-means training
> > *RBFNN Design using MATLAB's NEWRB
> > Retrain the created neural network
> > *Training Feed Forward Neural Networks
> > 3 threads for heath newrb overfitting -overtraining
> > Question Regarding RBF?
> > Neural Network -- Incremental Training
> > train rfb newrb
> > 2 threads for heath newrb -overfitting overtraining
> > See "*" above
What is the size of your output target matrix?