Bot-Breaking Backgammon Positions
by Douglas Zare
1 November 2012
When a backgammon bot disagrees with a human, it's usually right to bet that the bot is correct.
However, there are many backgammon positions the bots don't understand well, including quite simple ones.
All bots have weaknesses, but I'll show a few positions Gnu Backgammon misevaluates.
As usual, in all positions, Red bears off to the bottom left.
Here is a classic example: Rolling an outside prime home.
Red played an ultra-deep backgame with 15 checkers back, and hit a shot during the bearoff. Red is now rolling trying to roll home a prime. Red might roll 4-4 or 5-5 at the wrong time, and White may escape. Gnu Backgammon evaluations say Red wins 41%. However, with proper play Red is a strong favorite to close White out. White often wins after getting closed out, but I think Red should win 55-60% of the time with proper play.
Backgames like this are uncommon, and bots generally don't encounter them in self-play. (It is also very hard to run into them when both players play well. Sending the last few checkers back is usually a clear mistake.) Gnu's evaluations were better than I expected. It still isn't sure how much value White's home board points have, but it seems to have been trained to roll outside primes home. Some earlier bots would not recognize Red's position as winnable. I believe Gnu Backgammon doesn't play terribly in rollouts of this position, it's just that its absolute evaluations are off.
|The rest of this article (19.19 K) is premium content. Please subscribe below.
Article text Copyright © 1999-2018 Douglas Zare and GammonVillage Inc.