| |
BGonline.org Forums
Why do bots have an even/odd effect?
Posted By: MaX In Response To: Why do bots have an even/odd effect? (Tom Keith)
Date: Wednesday, 12 August 2009, at 3:27 p.m.
Tom wrote: Bot evaluation functions are typically set up with no input for "which side is on roll". Instead, the evaluation function assumes one side (say white) is on roll. When is it's black's turn, the bot simply inverts the colors of all the checkers (so that "white" is now on roll), feeds that to the evaluation function, and reverses the result that comes back.
Well, this seems the right way of doing it. No need to have a NN be trained for something we can compute easily with a subtraction. A NN with the "who's on roll" input will risk being inconsistent: the result of a given position with white on roll should be equal to the result of the dual position with black on roll. Even if you try to enforce this in the training, I doubt you can obtain something reasonable.
Tom wrote: This leads to an even/odd effect between levels of lookahead. Let's say it is white's turn. A 0-ply evaluation looks at the board from white's perspective. But a 1-ply evaluation looks ahead one roll before evaluating the position. So all the evaluations at 1-ply are from black's perspective.
This shouldn't be an issue. The 1ply equity of position P0 whith white on roll is the weighted average of white equity in the 21 subsequent positions obtained over the 21 different rolls played optimally (from white's point of view) according to 0ply. These equities are after white's moves, hence are with black on roll. But that's not an issue.
One thing I can figure out is the following: imagine a NN that systematically underesitmates the equity of the player on roll. A 0ply analysis with white on roll would then underestimate white's equity, but a 1ply analysis would overestimate white's equity. Of course this is not the case of gnubg. However, if it errs on a specific position and this position is N plies ahead and you analyse with exactly N plies, then you will "see" the error, while if you evaluate with more or less plies you may "avoid" the error.
Tom wrote:
While the even-odd effect might seem a bit disconcerting, it is not necessarily a bad thing. As long as the bot plays well, and gives accurate cube decisions, that's what is important.
I agree. But I've seen cases where 0/1/2/3/4ply evaluates a cube decision with wild swings that do not converge when moving towards higher plies. Also, the fact other bots do not exhibit the same beheviour makes me think that there's something wrong.
MaX.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.