Open motiwari opened 4 years ago
This would also allow you to lower p
and do majority voting for the same dataset. Right now p
is set very conservatively (small). p of 1/(kN1e3) is too large for certain corner cases like k = 50, N = 100 .... p = 1/(kN1e4) solves the problem
1 for drawing the sample, 1 for controlling the randomness in UCB. That way we can do multiple UCB runs for the same subsample