| |
BGonline.org Forums
Bot Comparison after 60/500 games... GNU odd/even ply cube effects
Posted By: neilkaz In Response To: Bot Comparison after 60/500 games. (MaX)
Date: Friday, 30 October 2009, at 2:27 p.m.
MaX "the idea of supervised training where 0ply is trained to mimic 2ply (instead of 1ply) is a bad idea, at the origin of the odd-even crap."
Neil replies... is this how GNU was trained? I am unaware of how GNU was trained to be honest, but this odd/even ply stuff can be severe. This effect is clearly less in other bots.
The chart shows a clear representation of the GNU odd/even ply oscillation effect. (The below is for casual readers here, as those who've been using GNU for years understand this)
3 ply almost always evaluatates that the side on roll has better winning chances than the what 2 and 4 ply think.
As a result, it cubes sooner and passes sooner on 3 ply. That effect spread across a vast universe of positions results in 3 ply often cubing too soon and passing too soon.
Readers here have often seen me stress that I want to use GNU 3 ply for basic holding game cubes, when using GNU. This is because 2 and 4 ply usually overvalue that type of position by a couple of percent GWC for the single anchor. 3 ply since it values the side on roll higher, usually lucks into a very good estimate of GWC, for holding games. However, as we see from the table, it will make too many errors in other positions.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.