Default parameters for net.trainParam and net.trainFcn - Neural Network Toolbox

I do not have access to the Matlab Neural Network toolbox at this time. What are the default parameters of net.trainParam and net.trainFcn if you create a network by the newff.m function, say, net = newff(feature_vector, output_vector, num_hidden_nodes), and if you train the output network with the train.m function, i.e. [net, tr, Y, E, Pf, Af] = train(net, feature_vector, output_vector)?

Answer 1

[Greg Heath]

 [inputs,targets] = simplefit_dataset;
 whos
 net = newff(inputs,targets,5)

 %   Name         Size            Bytes  Class     Attributes 
 %   inputs       1x94              752  double              
 %   targets      1x94              752  double  

 trainFcn = net.trainFcn                % 'trainlm' (Levenberg-Marquardt)
 mingrad =    net. trainParam.min_grad  % 1e-5
 mu =   net.trainparam.mu               % 1e-3 
 mudec =   net. trainParam.mu_dec       % 0.1
 muinc =   net.trainparam.mu_inc        % 10
 mumax =   net.trainParam. .mu_max      % 1e10

[Maxine]

Thanks for the reply. Is the momentum in the ANN training represented by mu, and the minimum learning rate, mingrad? Does the training of the ANN using the train.m function in Matlab start with a high learning rate and a low momentum? Then, over time, the learning rate is decreased and the momentum is increased or appropriately adjusted?

[Greg Heath]

Google

levenberg-marquardt backpropagation wiki

levenberg-marquardt matlab

[Berna Bulgurcu]

Good question. However there is no clear answer in google and matlab text. So is it possible to reply if you know the answer, Dear Greg?

[Berna Bulgurcu]

"net.trainParam.lr" shows learning rate. However, we can not see this code in the Levenberg-Marquardt matlab codes. Why?

[Greg Heath]

The obvious reason: TRAINLM does not use a learning rate. See the references to understand how the LM algorithm works.

[Berna Bulgurcu]

http://dali.feld.cvut.cz/ucebna/matlab/toolbox/nnet/trainlm.html. This page says that default learning rate is 0.01 for TRAINLM. Thank you for your answer.

Answer 2

[guirat nedjma]

changement de la valeur de net.trainParam.lr sur trainrp ne fait rien changer dans mes resultat