|
BGonline.org Forums
21$-66B-51P-42P-42 – Unlimited Games – 3-ply XG2 Rollout
Posted By: Timothy Chow In Response To: 21$-66B-51P-42P-42 – Unlimited Games – 3-ply XG2 Rollout (Ian Dunstan)
Date: Monday, 28 July 2014, at 5:26 p.m.
Ian wrote:
a 1296 gnubg rollout and 1296 XG rollout differ a LOT in their reported accuracy
I've noticed this too, and have wondered what's behind it. In general one should expect that a stronger bot should yield lower variance, because I think a perfect neural net should get the equity exactly right with every rollout trial. So I attributed the lower variance to XG's stronger net. (In fact, if I recall correctly, there was a noticeable difference between XG1 and XG2 in this regard.)
But that was for GNU 0.x rather than GNU 1.x. Are you using GNU 1.x? I haven't don't many rollouts with GNU 1.x and for those that I have done, I haven't looked closely at the variance to compare it with that of XG. The new GNU neural net should be close in strength to XG2, so I would expect that it should report similar results.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.