You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, isn't exp(-g/t) always going to be less than 1 as long as g/t is positive. Why is the min necessary? Also in the code:
I noticed that g/t is less than 1 most of the time so it never does the exponential.
The text was updated successfully, but these errors were encountered:
I believe that the code intends to impose a penalty of exp(-g/t) for programs that exceed the time budget on the score ( g/t > 1), else score is simply log2(n). The mathematical formulation is certainly wrong, but I am not sure if there is a simple way to express this logic.
The score is log2(n) * exp(-g/t) if the budget is exceeded.
So the cost function I was expecting was something like this:
So there is quite a big penalty for going over the time budget, with a bit of
a cliff, but still some reward. Originally I had min(1,exp(-g/t)/exp(-1)), but
then decided to penalise exceeding the budget a bit more - as @nichlzk says,
there isn't an elegant way of expressing that, and I seem not to have noticed
that the maths no longer follows the function.
Regardig @aaronlws95's point about g/t being less than 1 most of the time: that is because
the default prediction of n is extremely conservative. You'll want to try to make a better
prediction of n, so that you can get reasonably close to the g/t=1
Hi, isn't exp(-g/t) always going to be less than 1 as long as g/t is positive. Why is the min necessary? Also in the code:
I noticed that g/t is less than 1 most of the time so it never does the exponential.
The text was updated successfully, but these errors were encountered: