>From my understanding based on Joseph Heled's page, the NN training does not >try to make the %s match the true values, but rather tries to ensure that the >NN makes the best decisions. (This is a subtle difference, and it doesn't seem >to fundamentally change the question.)
I imagine that as the NN improves, so does the benchmark. How often/recently is the gnubg NN retrained? Do we think it can be improved much? Best, Aaron ________________________________ From: Joaquín Koifman <[email protected]> Sent: October 16, 2020 12:12 PM To: Ian Shaw <[email protected]> Cc: Aaron Tikuisis <[email protected]>; Øystein Schønning-Johansen <[email protected]>; [email protected] <[email protected]> Subject: Re: The status of gnubg? Attention : courriel externe | external email May I ask a couple of questions regarding NN training? >From my little understanding, I suppose there are 2 sources of errors during >an evaluation: a) the NN may be intrinsically unable (say, because of the type, number, etc of inputs) to "score" well, when compared with the true equity/%s (at least, the ones you want to replicate) of a position, or b) the benchmark against which the NN is being tested might not have the "true" equity/%s because, for example, the rollouts were done in 0-ply. Is there any way to know which of these two factors is limiting the most the improvement of gnubg's NN? I mean, do we need to improve the benchmark, do more training or change the NN altogether to improve the evaluation? Thanks
