BGonline.org Forums

Maybe there's a better way to measure takes?

Posted By: Fabrice Liardet
Date: Sunday, 18 October 2009, at 9:40 a.m.

In Response To: Maybe there's a better way to measure takes? (Phil Simborg)

IMHO you tapped into the main weakness of today's bots. Their evaluations are relatively dumb when it comes to cube (or recube) vig. If I recall correctly they use a fixed-valued vig of 60%. By the way, I think it is that 60% which is behind the famous 22% rule of thumb for the take point. If all recubes (to 4, to 8, to 16...) were perfectly efficient as in the so-called "continuous game", it is well-known that the take point would be 20%. If there were no recubes, the take point would be 25%. 22% is 60% of the way between 20% and 25%.

I am not sure whether it was GNU or Snowie which adapted the cube vig to certain class of positions. Maybe I recall reading that once in the GNU documentation. E.g. 70% for a straight race and 50% for a holding game with low racing chances. Still sounds like a rather primitive estimation. I don't know how XG does.

The right way for a bot to eval the cube vig would probably be to integrate it into the neural network ; but since the cube vig is very score-dependent that would require the neural network to be itself score-dependent. Zbot ?

As Timothy said, even when vigs can be reconstructed from the rollout results (but can one really separate the different cube, recube and re-recube vigs ?), since the rollout relies on the cube decisions made by the evals it is not sure that the results would be meaningful. One would indeed need the rollout to fork at each potential cube decision.

As a side note, the cube vig is even more important for double / no double decisions than the recube vig is for take / pass decisions, since the doubler's future cube vig is the exact reason for not doubling although one is in the doubling window. For that reason, while I have a reasonable confidence in what the bots say about take/pass decisions, I don't have such a confidence about double / no double decisions.

Messages In This Thread

• Maybe there's a better way to measure takes?
  Phil Simborg -- Sunday, 18 October 2009, at 1:58 a.m.
  • Maybe there's a better way to measure takes?
    Timothy Chow -- Sunday, 18 October 2009, at 3:15 a.m.
    • My thoughts on your responses
      Phil Simborg -- Sunday, 18 October 2009, at 2:53 p.m.
      • My thoughts on your responses
        Stick -- Sunday, 18 October 2009, at 3:01 p.m.
      • My thoughts on your responses
        Rich Munitz -- Sunday, 18 October 2009, at 7:07 p.m.
        • My thoughts on your responses
          Daniel Murphy -- Sunday, 18 October 2009, at 7:37 p.m.
          • My thoughts on your responses
            Rich Munitz -- Monday, 19 October 2009, at 3:29 a.m.
            • My thoughts on your responses
              Daniel Murphy -- Monday, 19 October 2009, at 5:24 a.m.
              • My thoughts on your responses
                Phil Simborg -- Monday, 19 October 2009, at 1:21 p.m.
                • My thoughts on your responses
                  Peter G. -- Monday, 19 October 2009, at 3:08 p.m.
  • Maybe there's a better way to measure takes?
    Matt Cohn-Geier -- Sunday, 18 October 2009, at 3:36 a.m.
  • Maybe there's a better way to measure takes?
    Andreas -- Sunday, 18 October 2009, at 8:36 a.m.
  • Maybe there's a better way to measure takes?
    Fabrice Liardet -- Sunday, 18 October 2009, at 9:40 a.m.
    • Maybe there's a better way to measure takes?
      Timothy Chow -- Sunday, 18 October 2009, at 2:22 p.m.
      • Maybe there's a better way to measure takes?
        Fabrice Liardet -- Sunday, 18 October 2009, at 3:58 p.m.
  • Maybe there's a better way to measure takes?
    Stick -- Sunday, 18 October 2009, at 1:15 p.m.
  • Maybe there's a better way to measure takes?
    Fabrice Liardet -- Sunday, 18 October 2009, at 1:37 p.m.
  • Maybe there's a better way to measure takes?
    Ian Shaw -- Sunday, 18 October 2009, at 4:12 p.m.
    • Maybe there's a better way to measure takes?
      MaX -- Thursday, 22 October 2009, at 12:31 p.m.
  • Maybe there's a better way to measure takes?
    NJ -- Sunday, 18 October 2009, at 5:14 p.m.
  • Maybe there's a better way to measure takes?
    Keene -- Monday, 19 October 2009, at 4:33 p.m.
  • Maybe there's a better way to measure takes?
    Daniel Murphy -- Monday, 19 October 2009, at 6:14 p.m.