grf-labs / policytree

Policy learning via doubly robust empirical welfare maximization over trees
https://grf-labs.github.io/policytree/
MIT License
74 stars 15 forks source link

Constrained optimal policy #93

Closed hhsievertsen closed 3 years ago

hhsievertsen commented 3 years ago

Hi team,

thanks again for developing these tools and being so responsive. Is there an easy way to find an optimal policy that is constrained to treating at most N individuals (i.e. I have a fixed budget)?

thanks, Hans

erikcs commented 3 years ago

Hi @hhsievertsen, policytree only does unconstrained search at the moment.

You can however shift the objective function by the average cost of treatment C by subtracting it from the doubly robust scores, see section 5.1 in https://arxiv.org/pdf/1702.02896.pdf for an example (the replication code is here.

(the supplemental material to https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA13288 has some MATLAB scripts that uses integer programming to find policies with constraints, they are not tree based however).

hhsievertsen commented 3 years ago

Hi @erikcs, perfect. That works. Thank you!