Silicon Highlands Backgammon review

Here’s an informal review of Silicon Highlands Backgammon as of 1996 by Peter Ozanne, from the rec.games.backgammon newsgroup, along with the program author’s reply. I have lightly edited the postings.

Subject:      Re: Genetic Backgammon from Silicon Highlands
From:         Peter Ozanne <ozanne@iol.ie>
Organization: Ireland On-Line
Date:         Sun, 16 Jun 1996 18:52:51 +0100

Hi Bob!

Well done in creating your fun game, and good luck with the marketing! I shall, however, now have to be a little cruel to be kind; most humans - even most of those who play bg - don't play all that well, so your game will do quite well against the casual player. We, that is oz and gnome on fibs, just tried it and, I'm sorry to say, gave it a good thrashing. I like the graphics - it moves well, but plays rather weak (perhaps equivalent to 1400-1500 fibs rating) and has the added edge that it doesn't respect the Crawford Rule, viz.; at 2-1 I won a Gammon with the cube on 2, taking me to match-point at 6-1; for the first game at match-point, the cube is off-limits. Your game ignores this and doubles, thus being able to win 2, 4, or even 6 points, whereas the leader can only win 1. This makes the leader's lead rather insignificant, so about 30 years ago the Crawford Rule was invented, saying you can't double for the first game at match-point. If the trailer wins this "crawford game" he or she will usually double at the first opportunity in all subsequent games. Thus the leader has at least some reward for those hard-earned points, otherwise it would be just too easy too catch up and win the match.

By the way, does your game always get the first roll, or was I just unlucky?

I'm sure I'll use it again, if only to sharpen my technique against relative beginners (and yes, I was playing level 7!). I'm sure a lot of newcomers to the game will have great fun with it; I believe it's pretty hard to create a BG program of master level: the most successful ones seem to be developed using neural net, such as TD-Gammon, and Jellyfish and Loner. JF is available to buy, and both JF and M/Loner play on FIBS and GamesGrid. Anyway, many thanks for the free sample, and every success and happiness in the future!

See you on FIBS or GamesGrid,

Peter Ozanne (oz).

Subject:      Re: Genetic Backgammon from Silicon Highlands
From:         Bob Landwehr
Organization: MicroLith
Date:         Thu, 27 Jun 1996 22:29:58 GMT

Thanks, Peter, for the thorough response, particularly the excellent explanation of the Crawford rule. I'm embarrassed to admit that I didn't even know of that rule when designing the program, but will certainly include it in the first major update. Meanwhile, the program does not consider the match score at all when evaluating whether to double or accept a double, whereas you may, so any advantage here is with the human player. (No reason for you to observe the Crawford rule when the program won't!)

As you've probably noticed by now, you were just unlucky in never getting the first roll. (You may have also noticed that doubles are sometimes rolled on the first roll--soon to be corrected.)

You are right about the difficulties in creating a master player. Backgammon is a marvelously subtle game which poses delightful challenges for the programmer. Neural nets are probably the best approach in the long run. I was surprised at first that my genetic model did as well as it did. However, after a few years practice I can now beat it most of the time, the stronger programs you mention have come along, and I've had time to think of several ways to improve the model, so it's about time to go back to the computer and cook up something better.

Or do we already have enough expert backgammon programs? Would a moderately strong program with more bells and whistles--maybe improve the sound effects, teach it to play Acey-Deucey or whatever--be more valuable than one that can beat you, me, and Jellyfish? Comments, anyone?