Why can't I use the mae error with the Levenber-Marquardt algorithm?

Hi, I'm training a neural network using a script I got using the matlab tool on neural networks.In particular I am using a timedelaynetwork for the prediction of a historical power series, I modified the network by inserting two hidden layers, one with a logsig activation function and one with a tansig activation function.I am using is the levenberg-marquardt, inserting the mae as a performance function, the message in the figure appears in the command window.

Why can't I use the mae with the trainlm?

Also, I would like to ask you, in your opinion is the architecture and type of network I am using to make the power prediction correct? or could it be improved in some way?

Answer 1

[Matt J]

Why can't I use the mae with the trainlm?

Just a guess, but Levenberg-Marquardt presumes that a Jacobian can be computed at the optimum parameter selection. In the ideal scenario where the optimal MAE=0, the Jacobian would fail to exist, due to the non-differentiability of at .

[Giuseppe D'Amico]

clear, thank you.

What could I use?

[Matt J]

Couldn't you just use trainNetwork, say with its default stochastic gradient descent algorithm?

[Giuseppe D'Amico]

I have never used it, would it be okay to use the trainNetwork function to train a network needed to predict a power time series?