%%%
%%% File "train_net.m": neural network training and generation  
%%%

% Caricare P e T;

P = Psim(1:round(size(Psim,1)/3),:)';
T = Tsim(1:round(size(Psim,1)/3),1)';

%%%
%%% Neural network parameters
%%%

Si = 4;  % Number of neurons in the input layer
Sh = 8;  % Number of neurons in the hidden layer
So = 1;  % Number of neurons in the output layer
         % It is equal to the rows of the matrix T

TFi = 'tansig'; % Sigmoidal tangent activation function
TFh = 'tansig';
TFo = 'purelin'; % Linear activation function

%BTF = 'traingdx'; % Function for the training of the  
                   % backpropagation NN, default
BTF = 'trainlm';   % Levenberg-Marquardt backpropagation 

BLF = 'learngdm';  % Backpropagation function 
                   % weight/bias, default

PF  = 'mse';      % Performance function Mean Square Error
                  % default

PR  = minmax(P);  % Equal to: [min(P')' , max(P')'], it
                  % determines minimal and maximal values of 
                  % inputs and output

val.P  = Psim(round(size(Psim,1)/3)+1:2*round(size(Psim,1)/3),:)';   
                                                 % validation data                  
val.T  = Tsim(round(size(Psim,1)/3)+1:2*round(size(Psim,1)/3),:)';
test.P = Psim(2*round(size(Psim,1)/3)+1:end,:)'; % test data  
test.T = Tsim(2*round(size(Psim,1)/3)+1:end,:)';

%net = newff(P,T,[Si Sh So],{TFi TFh TFo},BTF,BLF,PF);
                                            % Note: it generates a NN
                                            % with 4 layers!!!
net = newff(P,T,[Si Sh],{TFi TFh TFo},BTF,BLF,PF);                                            

%%%
%%% Parameters for the NN training
%%%

net.trainParam.epochs = 300;    % Number of epocs
net.trainParam.goal   = 1e-4;   % Value of the final error
net.trainParam.show   = 1;      % Show the plot after 1 epoch
net.trainParam.lr     = 0.05;   % Learning rate for trainlm function
net.trainParam.mc     = 0.9;    % Momentum constant: gradient value 
                                % during the training phase: if 0 -> 
                                % weights are changed only on the basis 
                                % of the gradient; if 1 -> gradient
                                % function is completely neglected

net = train(net,P,T,[],[],val,test); % training function

Ts = 0.05;   % Sampling time
             % NOTE: it should be equal to the sampling time used
             % for collecting the matrices Tsim and Psim!!!

net.sampleTime = Ts;             
             
gensim(net,Ts); % It creates the neural network in Simulink

return