On 20/03/2014 09:44, Luisa wrote:
When I train the MLP classifier, is the root mean
square error (RMSE)
output associated with the last training epoch or is it a mean of the
RMSE in all epochs??
The RMSE shown in WEKA's classifier output area is always calculated
once the model has been fully trained. Thus, in the case of the MLP,
once all epochs have been completed, the evaluation data will be
processed and RMSE, etc., will be calculated.
And lastly, is there a way to access the RMSE per
epoch, so that I could
see it's evolution?
You can monitor the MSE per epoch in the MultiLayerPerceptron GUI (you
need to configure the classifier to show the GUI). Note that
MultiLayerPerceptron performs stochastic gradient descent, not batch
gradient descent, so the network parameters are updated after each
training instance has been processed and the sum of squared errors is
accumulated simultaneously. (I always thought it performs batch
optimization, but I just looked at the code and it appears to perform
stochastic gradient descent.)