| |
BGonline.org Forums
72-game rollouts in XG
Posted By: Timothy Chow In Response To: 72-game rollouts in XG (eXtreme Gammon)
Date: Tuesday, 3 May 2011, at 3:27 p.m.
Xavier wrote:
I did not see much value in rotating on the 2nd roll.
How much gain will it gives?
This is a good question. I don't have a definitive answer, but here's one reason my intuition suggests that there is some benefit. Say we're rolling out a cube decision. By rotating the first roll, we're making sure we "explicitly" examine the effect of each possible roll. By rotating the second roll, we're making sure we "explicitly" examine the effect of each possible response to the first roll. If we rely only on VR, then we're only "implicitly" taking into account the different possible responses.
Experience seems to show that, as far as evaluations go, there is a "knee in the curve" at 3 ply. That is, we seem to gain significantly in correctness by going from 2 ply to 3 ply. To me, that suggests that rotating the 2nd roll should produce a noticeable improvement in the correctness. The issue is not so much the confidence interval as the chances that the bot will misassess the truth if it does not explicitly consider every possible 2nd roll.
This is of course very fuzzy, but on the other hand, it seems to me that rotating the 2nd roll can't hurt (as long as you rotate in a random order, as you said).
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.