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
39
forks
source link
(Idea) feature: proportional accept rate during all phases #324
So far, the adoption rate in the replay buffer has been linear based on maximum_episodes, but this means that the earliest episodes will be selected many times before the buffer is filled.
Even if the diversity in each batch will be decreased a little, it would be better to use a weight proportional to the number of current episodes so that the earliest episodes are less likely to be selected.
ikki407
commented
1 year ago
So far, the adoption rate in the replay buffer has been linear based on
maximum_episodes, but this means that the earliest episodes will be selected many times before the buffer is filled.Even if the diversity in each batch will be decreased a little, it would be better to use a weight proportional to the number of current episodes so that the earliest episodes are less likely to be selected.