It is possible to create a dynamic model using radial basis function in the matlab tool box?

I want to create a net that uses radial basis function in a NARX model: inputs contain delays and feed back. Something like this http://fr.mathworks.com/help/nnet/ug/design-time-series-narx-feedback-neural-networks.html but using a radial basis function.

Answer 1

[Greg Heath]

 close all, clear all, clc, plt=0
[ X, T ]                   =  simpleseries_dataset ;
net                        = narxnet(0:1,[],1);
net.layers{1}.transferFcn  =  'radbas';
net.divideFcn              = 'divideblock';
[Xs,Xi,Ai,Ts]              = preparets(net,X,{},T);
xs                         = cell2mat(Xs);
ts                         = cell2mat(Ts);

 plt = plt+1, figure(plt)
hold on
plot(xs(1,:),'LineWidth',2)
plot(ts,'r')
MSE00s              = mean(var(ts,1))
net.trainParam.goal = 0.01*MSE00s;
Ntrials = 10

 rng('default')
for i= 1:Ntrials
   net                  = configure(net,Xs,Ts);
  [ net tr Ys Es Xf Af] = train(net,Xs,Ts,Xi,Ai);
  R2s(i)                = 1-mse(Es)/MSE00s;  
end
plt = plt+1, figure(plt)
hold on
plot(ts,'LineWidth',2)
plot(cell2mat(Ys),'r--')
R2s = R2s

% R2s = [ 0.99548 0.99556 0.99997 0.99813 0.99528 % 0.99528 0.99973 0.99534 0.9925 0.99085 ]

Hope this helps

Thank you for formally accepting my answer

Greg

[Platon]

Thank you I'll try it

Answer 2

[Greg Heath]

Unfortunately, there are no options in the NNTBX that will yield this.

Either modify MATLAB functions or find non-MATLAB code.

[Platon]

Modify MATLAB function? you mean something like 1) creating a static radial basis function net 2) train it (open loop), then 3) using a close loop it is possible to simulate it as it was dynamic?

[Greg Heath]

When I wrote that I didn't think it out because I didn't think you would seriously consider doing that rather than just use the MLP version.

You would probably have to transfer the open loop weights and transfer functions to a timeseries function.

You could try to use radbas in a time series function. However, I'm not sure what would happen.

Nevertheless, since it is fast and easy, I would try it 1st just to see what happens.

I recall using radbas in one or more newff designs. You might want to search in NEWSGROUP or comp.ai.neural-nets.

[Edna Plazas García]

Hi Greg Heath

I've been looking at your code and it seems very good, but I have two question, the firts I have some data from a plant already taken, how do I call this code and serve this code with that data? attached a .mat with the data(are 2001 data), the first row is U and the second row is Y. the second is how do I validate data elsewhere in this simulation? as I use the function sim ?. Thanks you.

Answer 3

[Edna Plazas García]

Hi Greg Heath

I've been looking at your code and it seems very good, but I have a question, I have some data from a plant already taken, how do I call this code and serve this code with that data? attached a .txt with the data(are 2001 data), the first row is U and the second row is Y.

[Greg Heath]

The attachment is missing. Either try again or email it to me *.txt or *.m.

 U = input, Y = target   ?

 [ Ys Xf Af] = net(Xs,Ts,Xi,Ai);