| |
BGonline.org Forums
Cubeful equities: how to derive from cubeless?
Posted By: NJ In Response To: Cubeful equities: how to derive from cubeless? (Bob Koca)
Date: Thursday, 7 January 2010, at 9:04 a.m.
I do believe that Janowski's idea works, it's just that the correct formula is just slightly different than what was written in that paper. So to revise my statement, cubeless + _correct_ Janowski formula + lookahead is probably more accurate than a cubeful neural net. By the way, I have reason to believe that GNU does it "the right way" and they don't use the formula in that paper.
After thinking about it just now, I am more convinced than ever that a cubeful neural net wouldn't work. Take the case of a cubeful neural net trained for the score -9 -9. As part of the inputs, it would need the doubling cube value, which could be one of 7 values (centered,2,4,8, opponent 2,4,8). Now think about how you would train this neural net for a single board position. You would first need to train 7 positions instead of 1, because you would need to train it at each cube value. But what you are actually doing here is training the MET into the neural net.
If you already have a MET with exact values, it's definitely more accurate to use the MET directly instead of letting the neural net learn the MET through examples. It's sort of like using a multiplication table to compute 9x6=54 versus using a neural net's fuzzy logic to get 9x6=54.2.
Having said that, I bet that a cubeful neural net would work for a score like -3 -3, but the farther out you get the worse the net would be.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.