You are here

Closed loop forecast prediction


I m using a simple sinusoid curve as a dataset
Math.Sin(i * 2 * Math.PI / 200); for i=1 to i=40 000

The network is a simple feed forward network.

I ve no problem training the NN and get to a decent error rate (0.0001).
My issue is that when i do closed loop prediction i get a flat response after a few steps (the output predicted value remains constant).

I did all kind of experiments on the window input (from 1 to 200 input), and on the size of the NN or the error rate. It does not seem to help....
Activation function is tanh but i also tried normalizing between 0 and 1 and using sigmoid as activation.

can you please help me understand what i'm going wrong ?


Neural Network Forums: 
jeffheaton's picture

The error rate reflects a non-closed loop prediction. So your neural network should be recognizing the signals that the sinusoid curve changes direction and begins to dip. It should also recognize that it is in an upward or downward trend.

I am guessing that the cycles are either compressed or expanded and that is preventing the NN from recognizing what part of the cycle you are on. I would look at how big the input window is. I would also look at a segment of the curve from training compared with a segment from the closed loop and see how similar they look.

Theme by Danetsoft and Danang Probo Sayekti inspired by Maksimer