| |
BGonline.org Forums
Why do bots have an even/odd effect?
Posted By: playBunny In Response To: Why do bots have an even/odd effect? (Tom Keith)
Date: Wednesday, 12 August 2009, at 7:51 p.m.
Tom Keith: ~~Bot evaluation functions are typically set up with no input for "which side is on roll".
Yes, I've seen that in the GnuBg code. I've tried to form an explanation based on that fact but haven't been able to without supposing incorrect training.
Consider a position with strong contrast between the two sides, say white just finishing bearing in against a black anchor on the ace point, black with a full home.
A 0-ply evaluation will examine the completed bear-in, greatly in favour of white. A 1-ply will look at black's 21 rolls by making each move (a table crunch or a checker escape), inverting the resulting boards and evaluating each as if it were white in the unfavourable position.
I'm still not sure why there would be enough difference in these to create an odd/even play effect. It's true that these are markedly different positions but I would imagine that in training the nets for one, the other would be encountered and trained for to an equal degree as you can't play one without the other. Thus the equities should be complementary.
Perhaps training doesn't work that way? I admit that I don't know how training works. I thought it did it by playing games and training on each position that arose.
Tom Keith: ~~GnuBG was trained to make its 0-ply evaluation reproduce as closely as possible the results of it's 2-ply evaluation. [...] GnuBG was specifically trained to give accurate absolute equities in cube oriented situations.
I haven't a clue about this but it does say that training is a lot more involved than I imagined. It's going to take a while before I understand this. Thanks for the info, Tom. :-)
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.