Posted By: Nack Ballard In Response To: Here's what's wrong with the RO's posted so far. (neilkaz)
Date: Wednesday, 20 January 2010, at 5:02 p.m.
In Response To: Here's what's wrong with the RO's posted so far. (neilkaz)
I totally agree. (I'm glad I opened this -- I almost bypassed the thread.) In further support of your point, Snowie also evals your post 61-61 position as a .91 take.
I discovered this phenomenon myself several years ago when doing layered rollouts for simple 5pt and 7pt holding games.
["Layered rollouts" is Paul Magriel's term for testing future rolls or roll exchanges (usually one for each type of resulting position as 21x21 or even 21 is too many) and seeing what bots do with the checker plays and/or cube, and supplementing with rollouts from there as needed. Your 61-61 is a simple and straightforward example.]
In a similar vein, I noticed (back then) that Snowie's rollouts underestimated the leader's chances in a common midpoint + 8pt vs bar point showdown. In this case, it was due not so much to poor handling of the cube, but rather that Snowie wouldn't come down when it should with 54 and 53 (sometimes 52), leaving a shot on the 9pt or 10pt. The initial rollout results were biased against the leader because he wrongly paid later. I don't know but it seems likely to me that bots still have this showdown bias.
If bots are skewing simple positions, what chance do they have to get backgame and post-backgame positions even close to the mark? (That doesn't mean we shouldn't try anyway -- it's a key way forward.)
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.