How can I get the internal structure of idNeuralNetwork (Regressio​nNeuralNet​work) used in NLAR

I have generated the identification of an error from the function sys = nlarx(z,R,NL), where NL=idNeural Network and I would like to know the internal structure (what operations are carried out and with what weights). Thanks!!

Answer 1

[Shantanu Dixit]

Hi Daniel,

If I understood the query correctly, you're creating a nonlinear ARX model using 'nlarx' with a neural network nonlinearity ('idNeuralNetwork') and would like to examine its internal structure, including the operations and weights.

The neural network component in the model processes the regressors through multiple layers with specified activation functions. For example:

NL = idNeuralNetwork([10 5 2], ["relu", "tanh", "swish"])

After creating a nonlinear ARX model using 'nlarx' : sys = nlarx(z,R,NL), you can access the weights and architecture through 'Network' property of idNeuralNetwork as follows:

nn = sys.Nonlinearity;
disp(nn.Network.Parameters.Learnables(1)) 

% Name: "fc1_Weights"
% Value: ...
% Free: ...

% Parameters contains the learnable hyperparameters and initial hyperparameter values used by the network
% Learnables — Vector of tunable parameters that represent the weights and biases for the network

You can also refer to the 'idNeuralNetwork' documentation: https://www.mathworks.com/help/ident/ref/idneuralnetwork.html for more information.

Hope this helps!

[Daniel]

Thanks!! And how can I know the intern opperations that are carried out with those weigths and bias in order to get the final estimation of the NN?