Open heidmic opened 1 year ago
There is one point left open are rules with experience smaller than z times dimensionality reliable at all? (e.g. z==0.5*d or z==5 assuming no relations to dim)
. This basically asks whether we need a lower bound as well. Maybe something that decreases based on a function (probably nonlinearly). But it might also be detrimental as it strongly discourages specific rules. To implement this in a good way, we would probably need to do something like estimating the "relevant" experience based on some kind of neighbourhood estimation
We want to add more variants to https://github.com/heidmic/suprb/blob/master/suprb/solution/mixing_model.py
Current mixing is making a model prediction based on all matching classifier's predictions times their inverse errors divided by their experiences (weighted sum).
What we want to test:
y times dimensionality
are equivalent in their weight to rules of experience> z times dimensionality
(assuming identical errors)