|
BGonline.org Forums
Results-Based PR
Posted By: Maik Stiebler In Response To: Results-Based PR (Christian Plenz)
Date: Monday, 2 February 2015, at 8:55 a.m.
I think this converges with luck% ...
Technically, it doesn't, unless your evaluations are actually perfect. As a thought experiment, stage a bot match where Bot A is the imperfect bot that also does the later analysis, and Bot B is a future generation bot which is stronger than Bot A - as we know, Bot A is not perfect, so it is not unreasonable to assume that such a bot exists or will exist.
Obviously Bot A plays perfect according to analysis, and Bot B does not, so by concventional mwc analysis we will converge to the wrong result that A is favorite. That won't happen with VR, which is unbiased by construction.
If that difference matters in practice is another story... ok, currently in the news is the Mochy game where he played a super backgame with a grotesque ER>30 (according to gnubg), apparently (see Daniel Murphy's comments, I haven't had time yet to look at the thing myself) to a considerable extent due to the bot misunderstanding what was happening. If that's true, a luck based analysis would probably have worked much better for that game - I mean probably it wouldn't have produced useful information either, but it wouldn't have punished Mochy so hard.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.