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

BGonline.org Forums

An interactive backgammon training website.

Posted By: Michael Petch
Date: Saturday, 4 January 2014, at 2:56 a.m.

In Response To: An interactive backgammon training website. (Mislav Kovacic)

On the RGB newsgroup I was asked about what improvements there were in the 1.x releases. One of them seems to be a reduction in the even/odd ply effects. 0.90 neural nets were known to have poorer cubing than the even plies. (3 ply cube was weaker than 2 ply cube). The link in my post gives you an idea of what the Depreli study thinks about the cube play.

On 11/08/2013 7:20 PM, badgolferman wrote:

> What is better about this version?

Depends on what version you were using previously.If you were using a release prior to about 2010 there are some speed improvements, but as of this year we are using a newly trained neural net.

Since beta 1.00 we have been using a stronger neural net that seems to have mitigated some even/odd ply effect. The Depreli study ( http://www.bkgm.com/articles/Keith/DepreliBotComparison/ ) is one benchmark that is used to compare bots. For comparison, the 2012 Depreli update can be reasonably compared with the numbers that I have produced with the older neural network and the new (1.0x) neural network. That new data is here: http://www.capp-sysware.com/downloads/gnubg/gnubg_prelim_results-1_00.txt

One may note that 2ply (WorldClass) v1.00 outperformed (for the Depreli positions) v0.90 4ply. The biggest changes are cube actions for Grandmaster, and cube handling overall.

Caveat: These studies focus on positions that come up in bot-versus-bot play. Human-versus-human play might yield different types of positions.

Xavier, Frank Berger, Michael Depreli and I have done some preliminary work for the next Depreli study, and I'm under the impression that the earliest such a study would be forthcoming would be in 2014.

Beyond that there have been a number of bug fixes and improvements. For those who automate analysis via the Python interface that interface has been improved.


Messages In This Thread


Post Response

Your Name:
Your E-Mail Address:

If necessary, enter your password below:




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

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