Bug in nonstationary error estimation

[greenTara]

The time value that goes into the method .eval (on a qmm) when calculating the difference between the estimated transition matrix and the model transition matrix should be the time at the center of mass of the weight distribution. When current_time = window_size + k * self.shift -1

For the naive algorithm the center of mass is: dk =int((window_size-1)/2) estimation_time = current_time - dk

For the bayes algorithm this is: w = window_size s = shift dk = int(w -( s+1)/2 - w * math.pow(r, k+1)/2 ) estimation_time = current_time - dk

[greenTara]

See also the comments on issue #4

[alexlafleur]

When I change the formulas, the plots look like this:

Old script:

New script:

[greenTara]

ok, that's more like the expected result.

[greenTara]

This appears to be resolved.