Training a NarX net with multiple datasets

Hi, i´ve been doing some experimentation training NarX networks with multiple datasets from the same model. Supposedly each time the net is trained with a dataset the weighs are modified to find an optimum. However, if you train that same network with the modified weighs with other dataset, i should change them again to fit the new dataset, right?

The thing is, that after training the same network with 6 different datasets from the same model and the validating with other datasets, the net had matched the model dynamic perfectly.

My question is, is there a name for this specific way of training a neural network or was it just luck?

Thank you beforehand

Answer 1

[Greg Heath]

Typically, training with a sequence of multiple datasets tends to cause the net to forget the salient characteristics of former training subsets. However, if the different subsets can be assumed to be random choices of realizations from the same model, you should be OK.

The salient characteristics should probably include min, median, mean, std, max and significant auto and/or cross correlation delays.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Luis Ignacio Ruiz]

Thank you very much for the answer. How can I see the salient characteristics?

[Greg Heath]

Other than staring at a plot, you have to compute them from the data.