Bot Comparison Final Table
Posted By: Ian Shaw In Response To: Bot Comparison Final Table (Michael Depreli)
Date: Monday, 26 April 2010, at 11:22 p.m.
In Response To: Bot Comparison Final Table (Michael Depreli)
Thanks, Michael, to you and your team for the you've effort put in.
Of great interest is gnubg's cube action. I see two sources of potential error. The first is a evaluation of the absolute value of the position. This is down to the strength of the raw nn. The second is how the cubeless numbers are converted into cube actions. If you look at the results for 2- and 3-ply there are some interesting features.
Gnubg 2-ply chequer play is very slightly better than Snowie 3-ply, but the cube results are worse. This suggests that gnubg is not making best use of its cubeless evaluations.
Secondly, gnubg 3-ply is clearly passing far too often. It makes the fewest bad Takes but the most bad Passes. Compare to 2-ply, where gnubg is doing the opposite: gnubg makes more bad Takes and less bad Passes than XG or Snowie. I interpret this as evidence that gnubg is under-valuing the position of the player on roll (because the cube action is evaluated for the doubler and simply inverted for the taker.)
Again, looking at the double/no double errors, gnubg seems to err by missing cubes rather then being to aggressive.
Also of note is that three-ply cube is weakest overall even though it has been shown to be better than 2-ply at evaluating holding games, where 2-ply makes bad takes as the holder. So it must be really messing up other actions.
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.