DeNA / HandyRL

HandyRL is a handy and simple framework based on Python and PyTorch for distributed reinforcement learning that is applicable to your own environments.
MIT License
282 stars 42 forks source link

(Outputted results will be changed) output whole results #354

Open YuriCat opened 7 months ago

YuriCat commented 7 months ago

If results for evaluation episodes that were not completed by the end of the epoch are truncated, not only will the number of results displayed be reduced, but there will be some bias on the results by the fact that longer episodes are more likely to be ignored.

While there is a downside to the results displayed not matching the epoch, the concern about results being discarded is more significant in practical development.