[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Calculating Cube Efficiency from Rollouts/Evaluations

Posted By: Rick Janowski
Date: Saturday, 12 January 2013, at 8:23 p.m.

In Response To: Calculating Cube Efficiency from Rollouts/Evaluations (Bob koca)

Interesting question. I think 1-ply cubeful equity evaluations could potentially be improved in 3 ways:

(1) Improved estimates of cubeless winning probability.

(2) Improved estimating of gammon/backgammon proportions (ie, average cubeless Win and Loss values).

(3) Improved estimating of cube efficiency.

XG’s 1-ply already does a remarkably fine job in (2) consistent yielding estimates of W and L very close to those obtained by XGR++ evaluations and 3-ply rollouts. No significant need for improvement here in my view.

In (1) XG’s 1-ply estimates of cubeless winning probability are generally fairly good but not always. I think it often misevaluates the strength of backgames, various types of busted board positions, and perhaps other highly complex positions. Still room for improvement here perhaps, especially in backgames.

In (3) XG’s use of a single cube efficiency factor at the 1-ply stage (as opposed to attempting to input separate x1 and x2 values) generally works very well in practice. Subsequent ply-levels effectively get a handle on the actual short-term volatility of the position, in a way akin to investigating the range and nature of potential market-losers. XGR+ and especially XGR++ also manage to get a reasonable handle of the more long-term likely volatility of the underdog (the taker typically). Using separate x1 and x2 values at the 1-ply stage would be quite complex and prone to potential problems, because the factors tend to vary considerably with winning probability. Although I think XG’s approach seems to work well in practice, there might be scope in refining estimates of x for various game types, but generally I think only a little tweeking might be needed. Games with busted structures tend to have low cube efficiencies so perhaps an initial x-value of 0.5 would be more suitable than the default 0.705. High-anchor holding games, 22-pnt holding games and mutual holding games might benefit from a slightly raised x-value, but practically very marginal benefits might be outweighed by other programming complications. Potentially room for improvement here but less of an issue compared to (1) estimation of cubeless winning chances, in my opinion.

Messages In This Thread

 

Post Response
Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:
 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.