| |
BGonline.org Forums
Which play is better based upon rollouts
Posted By: Timothy Chow In Response To: Which play is better based upon rollouts (David Rockwell)
Date: Tuesday, 27 April 2010, at 2:41 a.m.
David Rockwell wrote:
The biggest problem with it is that EITHER side can end up playing a backgammon or something not handled well within the rollouts. Because of this, errors in the middle game and end game will tend to be offsetting and too small to hit the threshold of human interest.
This is vaguely plausible reasoning, but we really don't know that it is correct. It's certainly true that either side can end up playing poorly. But we don't know that the errors will cancel each other out enough to make it negligible.
I can imagine trying to test your argument, as follows. Come up with some crude, fast, but hopefully reasonably reliable way to check if a position is one that a bot will make an error on. We don't have to catch all errors, just a significant fraction of the big ones. Then during a large rollout, use this subroutine to estimate how many of these errors occur on one side versus the other. If the errors tend to balance out then you'll have some evidence for your claim.
When I lay this out explicitly, I think you can see the difficulties. It's hard to imagine coming up with a reliable recognition subroutine for "positions that a bot will screw up." Even if we came up with one, how would we know how good it was?
These difficulties just underline the fact that we don't have hard evidence that systematic errors will cancel out. In the absence of any solution, it makes good practical sense to assume that they will cancel out, and just hope for the best. But my point is that we shouldn't delude ourselves into thinking that this assumption is anything more than an assumption that is not backed by hard evidence.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.