My textbook/manual understanding of MLP is that the estimator should be the squared classification error or cost function, e.g.:
sum{i=1,N (1/2*(y_i-t_i)^2)}
where N = number of events, y_i=MLP output for event i, and t_i=target/truth value for event i (0 or 1)
I ask because my plots of estimatorHistTest and estimatorHistTrain produced by MethodMLP are not what I expect given this definition. My estimatorHistTest is consistently lower than estimatorHistTrain and the two have nearly identical distributions. Also, they hit a minimum around 100 epochs, but never come back to that minimum, later flattening out around 600 epochs at a higher value.
I am using the defaults of sequential backpropagation and random input event sampling each epoch. I’ve tried plotting the estimator using batch processing instead, but something is crashing (probably on my end).
Thanks in advance!
Malina Kirn
lowly grad
University of Maryland