| |
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: Tuesday, 18 August 2009, at 12:59 a.m.
Tom: Suppose A has a gap on his 6-point and let's say the true value of A's position is 0.63. At 0-ply, the NN will incorrectly report A's equity as 0.73
At 1-ply, the lookahead routine takes a roll for A and plays it. B is now on roll, and the position is evaluated from B's point of view. B doesn't have a 6-point gap so the NN correctly evaluates the position from B's point of view.
To see that error, wouldn't it suffice to put B on roll and do a 0-ply evaluation? The NN would evaluate it from the position of a player without a gap on the 6-point and return -0.63; correct, though negated, but not matching A's 0.73.
If this is the case then is such a fault because the training is asymmetric? I understand the idea that you wouldn't want the player on roll to be an input to the NN but I don't understand why all of the training positions aren't entered in both positive and negative forms. If it's fit that the NN understands one side of a position then it should surely have as good an understanding of the opponent's view?
Yet if the training is already done this way then what other causes are there for a position and its anti-position to evaluate significantly differently?
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.