GNU BackgammonThe official website, with source and binary downloads.
| Jay : game learning : backgammon : TD-inspired |
All the current programs play nearly perfectly. There are slight differences in playing strength, but they can be detected only by careful measurement. You can’t tell any difference in strength by playing against them, even if you are an expert. You could play a tournament of thousands of games between the programs and still not know which was stronger, because the effect of the dice would be bigger than the effect of skill differences.
So choose a program by features, not by strength.
Free and open-source, and with powerful analysis features. It’s called “GnuBg” for short. It is only moderately good in usability—not bad, but with room for improvement. I think it’s the only backgammon program most players need.
GNU Backgammon
The official website, with source and binary downloads.
As CPUs get more and more cores, the best search to use in a backgammon program tends to shift away from classical full-width search (universal in backgammon since the 1990’s but difficult to parallelize) and toward rollouts (embarrassingly parallel but much slower). Extreme Gammon is the only program so far to be written with this insight in mind: Some of its settings allow hybrid searches that include truncated rollouts. This may be what makes it minutely stronger than other current programs.
eXtreme Gammon
The official website of the commercial program.
By Olivier Egger, Snowie is the most-mentioned commercial backgammon program. Version 1.0 was commercially released in July 1998, and as of 2012 it is up to version 4.
Snowie
The official website, with a free trial version.
By Frank Berger, Bgblitz (variously capitalized) is a less-often mentioned commercial program. Like the others, it is extremely strong. It remains under active development as of 2012.
Bgblitz
The official website, advertising availability on smartphones.
The Michael Depreli Bot Comparisons by Tom Keith
Since all these programs play extremely strongly, it takes a close analysis to find the differences in strength. This study works by selecting those relatively few positions where the programs disagree on the best move and rolling them out to get (a better idea of) the truth. Data is included for program versions into 2012.
This program by Fredrik Dahl was the first strong commercial neural network program and still has mindshare. As far as I can tell it isn’t sold any more.
Fredrik Dahl wrote to me that JellyFish uses a home-grown temporal difference algorithm, rather than TD(lambda). He was understandably reluctant to release details about the program's backgammon features and network architecture, but you can check out his publications.
JellyFish Backgammon for Windows
The official JellyFish home page, with descriptions and no-longer-functional ordering information, and a free Lite version. The lastest info seems to be 2004.
Harald Wittman wrote this program. Mloner (and another version restricted to one-point matches, loner) maintained one of the highest rankings on the First Internet Backgammon Server for nearly a year as of April 1996, but has long since retired. I believe it never went commercial.