Primal Bound and Integral #109
Replies: 2 comments 5 replies
-
An idea to deal with Primal Integral not being defined as long as no solution are found could be to split the reward functions in two.
has_no_solution = ecole.reward.NSolutions() == 0
penalty = 1000
my_primal = min(penalty * has_no_solution + ecole.reward.PrimalIntegral()) @gasse @dchetelat what do you think? |
Beta Was this translation helpful? Give feedback.
-
Sharing the issues about using If we have the SCIP primal bounds:
And in Ecole we want to provide rewards functions as differential improvements, then we would return
Which completely erases the Should we
I think I like a) better. It's the solution of least effort and not using any conventions. Something like reward_functions = ecole.reward.lambda(lambda model: model.as_pyscipopt().getPrimalBound()) |
Beta Was this translation helpful? Give feedback.
-
We'd like to implement a primal integral reward function.
What's a good definition/ implementation for it?
Beta Was this translation helpful? Give feedback.
All reactions