search

takuseno / d3rlpy

An offline deep reinforcement learning library
https://takuseno.github.io/d3rlpy
MIT License
1.29k stars 230 forks source link

[REQUEST] Support for additional OPE algorithms #231

Open LXYTSOS opened 1 year ago

LXYTSOS commented 1 year ago

Support for additional OPE algorithms such as weighted importance sampling, doubly robust.

pathway commented 1 year ago

Gawd yes please

takuseno commented 1 year ago

@LXYTSOS Thank you for the issue! Those should definitely be supported for the practical usecases. I'll consider this feature in the following updates.

paulhilders commented 1 year ago

@takuseno This would be a very useful addition to an already amazing library! Do you maybe have an indication of when these features might be added?

joshuaspear commented 1 year ago

Hey - I have started creating an ope library for offline RL which includes an api for d3rlpy. It's not fully tested yet but I have included regression (weighted importance sampling). I have also included functionality to have ope methods as scorers when calling LearnableBase.fit in d3rlpy. However, this relies on an epoch callback functionality that I have implemented in the following PR (https://github.com/takuseno/d3rlpy/pull/286). Feedback and contributions are more than welcome :)

https://github.com/joshuaspear/offline_rl_ope