ajul
commented
7 years ago Current solution is to use a less stiff rectifier function than the exponential, specifically one that is reciprocal for negative values and linear for positive values. While it lacks the mathematical elegance of the exponential, this appears to pass thousands of random trials without issue.
MultiplicativeBalanceoccasionally:expto overflow, generally a few times per thousand random trials.While this is not common it is still concerning. I need to figure out why this is happening and fix it. Current conjecture is that the exponential handicap function can be very sensitive to moderate changes in the variables, leading to instability.
This also seems to only occur in the weighted case, which is also curious.