|
BGonline.org Forums
Which play is better based upon rollouts
Posted By: Daniel Murphy In Response To: Which play is better based upon rollouts (Timothy Chow)
Date: Monday, 26 April 2010, at 10:26 p.m.
Tim: Suppose, for example, that bots play deep backgames poorly. We already know that this is the case
Tim, I think you may be overstating the bot/backgame problem. I think bots play "normal" backgames quite well, generally. If by "deep" you mean only backgames with many-many checkers back and uncoordinated front positions, then I think I agree with Matt, that such positions are rare with best (not bot but best!) play, and have little effect on rollouts of normal positions. Where I do see some problem with Gnubg, of unknown importance, is in its evaluations in positions with several (but not many) checkers back -- games that could sometimes be profitably turned into deeper backgames, but usually not (so checker play should not be a problem, but cube decisions may be).
It's hard to get a handle on bots and extreme backgames. I can willfully develop immense backgames against Gnubg and take/beaver cube after cube that the bot would not double if only it knew that I knew how badly if it would play from there. But then I've only shown that I have positive equity against Gnubg. In such games arise many positions where I can count on the bot screwing up by making plays a human would never make. So it's hard to estimate from those play sessions the true equities of those positions and strategies in human/human play. However, my judgment as a player is that it can't be right to systematically skew towards backgames, so a "systematic skew away from plays that tend to lead towards backgames" is, well, good play ;-
By bots I mean Snowie and Gnubg. I look forward to seeing more of the serious testing of XG that Snowie and Gnubg have endured in position types they tend to misplay or misevaluate. For example, "deep" backgames, prime walking positions, 22-point holding games, and 20-point holding games.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.