When XG is wrong
Posted By: Timothy Chow In Response To: When XG is wrong (phil simborg)
Date: Monday, 31 March 2014, at 2:43 p.m.
In Response To: When XG is wrong (phil simborg)
A little while back, I visited BGO briefly for news about the Dual Duel, and saw a couple of posts that I sent email replies to. I figure I might as well post the information here for others to see.
Corbett's book was full of positions the bots got wrong and he proved he could beat the bots in those positions. One of the things I did to prove XG was better was to put 30 of his positions into XG and found that XG got 28 of them right and the other two, very close, and with rollouts, right. (Or as "right" as could be determined by the rollouts and the experts like Corbett.)
I haven't put all the positions from Corbett's book into XG but I did find one or two discrepancies. Who knows what is "right," but for example I recall that in the "Kitten and the Puddle" chapter, a 4-ply/4-ply XG rollout disagreed with Corbett's verdict—if not in the main position, in at least one of the secondary positions.
Now my question is, and this is for you "experts" or very top players: where don't you trust XG? … But are they "types" of positions or games where you feel you must always roll them out? Are there types of positions or types of games where you are not sure you even trust the rollouts?
Here are some standard examples.
Superbackgames/prime rolling. See also this example from practical play.
Noncontact races (!)
The last three of these might be fixable with better neural-net training, but the superbackgames/prime rolling will require a different algorithmic strategy entirely. As I've said before, my hunch is that a combination of dynamic programming and neural nets would be the way to go.
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.