Inquiry about some Details on building a Neural Network

Hello All,

I have the following questions regarding building a Neural Network in Matlab

1- Is there a way to add a new Gaussian function as an activation function (Transfer function) in the neuron 2- Is it possible to build a custom network with a custom temporal Difference back propagation algorithm instead of the normal back-propagation algorithm used ? .. Is it possible to update the existing back-propagation algorithm to deal with the temporal difference algorithm instead ?

Really appreciate your reply Thanks so much

Answer 1

[Greg Heath]

0. "in the neuron" ??? Do you mean " in a hidden neuron layer " ?

1. NEWRB automatically adds radbas neurons during training. Additional radbas neurons can be added by modifying net.IW{1,1}. However, the performance on previous data will probably be compromised. Therefore, net.b{2} and net.LW{2,1} will also have to be adjusted.

2. The only way to do this using existing MATLAB functions is to

 a. Retrain NEWRB from scratch 
 b. Move the weights to NEWGRNN (regression) or NEWPNN(classification)

OTHERWISE

3. MODIFY NEWRB SO THAT AN INITIAL CONFIGURATION CAN BE SPECIFIED BEFORE LEARNING

4. It looks like TDBP is not a NNTB option.

5. You can always modify MATLAB's source code.

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 2

[Karen]

Thanks so much Sir for your answer , Yes in the neuron I mean the hidden layer neurons.

- Would you please advise what NNTB stands for ?

- I can't get the point Move the weights to NEWGRNN (regression) or NEWPNN(classification) , would you please advise.

- The radial basis function in Matlab is a = exp(-(n.*n)); , while the gaussian function that I need is

that means that I should add my own transfer function to the NNet toolbox right or is there something that I don't understand ?

- I want to build a neural network as the one attached in the photo, The output layer is a Gaussian function.

- The network has on hidden layer and one stochastic layer , instead of updating the weights using the equations of the normal backpropagation , I want to use other equations (That of temporal difference) , Kindly note that the links between the stochastic layer and the output layer has no links ,Is it possible to build such a thing using Matlab?

Really appreciate your response , Thanks so much