search

facebookresearch / off-belief-learning

Implementation of the Off Belief Learning algorithm.
Other
45 stars 7 forks source link

Any suggestions for the hyperparameter aux_weight #9

Closed PaladinEE15 closed 2 years ago

PaladinEE15 commented 2 years ago

Hi Hengyuan,

There is a hyperparameter aux_weight in selfplay.py, and I guess that it indicates the weight of the auxiliary task loss term. However, I cannot find the recommended value for it. Any Suggestions?

Sincerely, PaladinEE15

hengyuan-hu commented 2 years ago

Depending on what you want. If your goal is to maximize the selfplay performance, the AUX is not necessary, i.e. 0 may work quite well. If you want to maximize cross-play scores over several seeds, then >= 0.25 is good. We use 0.25 in most experiments when AUX is involved.

On Tue, Jun 21, 2022 at 4:39 AM PaladinEE15 @.***> wrote:

Hi Hengyuan,

There is a hyperparameter aux_weight in selfplay.py, and I guess that it indicates the weight of the auxiliary task loss term. However, I cannot find the recommended value for it. Any Suggestions?

Sincerely, PaladinEE15

— Reply to this email directly, view it on GitHub https://github.com/facebookresearch/off-belief-learning/issues/9, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABECKZNYZ6TU4XP2DPYJU3DVQF5VJANCNFSM5ZLPSIHQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

PaladinEE15 commented 2 years ago

Thanks for your suggestions