Of course you need to weight every position with the probability it occurs in actual play. (not that we know how to effectively compute it).
I don't think 2009 threads are a good indication. We need something with the current net, which I think is better. -Joseph On Sun, 8 Dec 2019 at 09:33, Timothy Y. Chow <[email protected]> wrote: > On Sun, 8 Dec 2019, Joseph Heled wrote: > > Yes. But there is the question of how easy it is to "navigate" to those > > positions. can you reliably get to those positions against a bot and win > > from their ignorance?I have my doubts. > > I've done some experimentation of this sort, but rather than quote my own > experience, I'll point you to reports by people who have spent more time > on this than I have. BGOnline is having some server problems, but I think > these links should work: > > http://www.bgonline.org/forums/webbbs_config.pl?read=129969 > http://www.bgonline.org/forums/webbbs_config.pl?read=54183 > > The upshot is that GNU 2-ply (version 0.xx) or XG 3-ply (version 1) is > reportedly exploitable in this manner, if you play money games with an > unlimited cube. Stronger settings are reportedly harder to exploit, but > maybe not impossible. > > In any case, as I said earlier, in my opinion this line of investigation > is not really the most interesting one. IMO the closeness of a bot to > perfection should be measured not just by its performance from the > standard starting position, but from *any* position that could legally > arise in a game. The reason is that in practice, bots are used to analyze > positions that arise in actual games. It's obviously a logical error to > conclude that the bot's analysis of an arbitrary position is sound just > because it plays well from the starting position, but I'm surprised to see > how often I see people implicitly making this error. > > Tim > >
