BGonline.org Forums

GNU

Posted By: Mike Petch
Date: Sunday, 28 April 2013, at 4:44 a.m.

In Response To: GNU (Seth)

The next version will be using the newer neural net (v0.91). Such a release may now be pushed back to next weekend. I spent part of this week running the Depreli positions. The data used is close to what was used for XG's 2012 Depreli Study. That can be found here:

http://www.bkgm.com/articles/Keith/DepreliBotComparison/

The Depreli study is a benchmark designed to see how well the bots play against each other (not necessarily bot vs humans). Please see the notes at the bottom of that link for some additional comments on the matter.

If we use the Depreli data that I generated this week using World Class (2-ply), Grandmaster(3-ply), and 4-ply on both the older GNUBG (20121023 Windows Release) and the new test version (0.91), some general things I can say without posting any specific data is:

- New GNUBG World Class (2ply) outperforms 4ply on the older version - All three playing levels of the new version exceed (or are comparable) to the performance of the best results (4-ply) of the older version - The best improvement was for 3-ply (Grandmaster). It seems the cube actions at 3-ply are significantly better. GNUBG suffered greatly with an even/odd ply effect, and the effect may be less of an issue in the new version. - Overall, cube play was much better between the old and new version.

Compared to XG2 and XG1 in the 2012 chart at the link above, it seems the new GNUBG on 4ply may be about as strong as XG v2, 4-ply. GNUBG's GrandMaster (3ply) is about as strong as XG v2, 3-ply. GNUBG's 4ply is a reasonably close to the strength of XG v2, 3-ply red.

I think that Grandmaster (3-ply cube, 3-ply checker) on the newer version may be the best option when looking for reasonable strength at a reasonable speed.

Note: 4-ply look ahead in XG is equivalent to 3-ply in GNUBG, 3-ply in XG is equivalent to 2ply in GNUBG and so forth.

I'm not posting specific numbers as I still have to validate the results, but at this time the observations above seem to be accurate given what the Depreli study measures.

Kudos to Philippe Michel for doing the retraining of the neural net.

Messages In This Thread

• GNU
  Bill Patterson -- Monday, 22 April 2013, at 4:29 a.m.
  • GNU
    Mike Petch -- Monday, 22 April 2013, at 5:44 a.m.
    • GNU
      Seth -- Sunday, 28 April 2013, at 2:50 a.m.
      • GNU
        Mike Petch -- Sunday, 28 April 2013, at 4:44 a.m.
        • GNU
          Michael Depreli -- Sunday, 28 April 2013, at 11:46 a.m.
          • GNU
            Seth -- Sunday, 28 April 2013, at 10:05 p.m.
        • GNU
          Mike Petch -- Sunday, 28 April 2013, at 2:31 p.m.
          • GNU
            Smcrtorchs -- Monday, 29 April 2013, at 3:16 p.m.
            • GNU
              Mike Petch -- Monday, 29 April 2013, at 7:48 p.m.
        • GNU
          Timothy Chow -- Tuesday, 30 April 2013, at 2:59 a.m.
          • GNU
            Mike Petch -- Tuesday, 30 April 2013, at 3:50 p.m.
            • GNU
              Timothy Chow -- Thursday, 2 May 2013, at 11:35 p.m.
            • GNU
              playBunny -- Tuesday, 7 May 2013, at 3:49 p.m.
        • GNU- New Version ?
          Seth -- Thursday, 2 May 2013, at 3:05 p.m.
          • GNU- New Version ?
            Mike Petch -- Thursday, 2 May 2013, at 7:19 p.m.
            • GNU- New Version ?
              Seth -- Thursday, 2 May 2013, at 9:50 p.m.