Closed diditforlulz273 closed 2 years ago
Yes, I see, there seems to be some problem here. Let me double check whether I make a mistake.
I checked the code and the paper, the current implementation can be considered correct when β = 0.
In the paper, author says that :
and at the end of the paper, they says that:
So It's easy to see that when β = 0 ,
always equal to 1/2.
ok, I then guess it is not the first-line thing to fix(as it works correctly in the end), feel free to close this issue.
ok, I then guess it is not the first-line thing to fix(as it works correctly in the end), feel free to close this issue.
Okay, closing issue. Feel free to reopen it anytime.
Am I missing something or do you miss DCFR beta round multiplication here: https://github.com/bupticybee/TexasSolver/blob/02e52e4bb47473d2854626f8c19f26e698c73347/src/trainable/DiscountedCfrTrainable.cpp#L111
I mean, you use a pure beta, not iter_num^beta / (iter_num^beta + 1), making this correction for alpha_coef some lines before.