Backgammon went the way chess did, only quieter and earlier: the program plays better than any human. That is not an opinion but a measured fact. The best programs — eXtreme Gammon (XG), GNU Backgammon and Snowie — beat the strongest people over the long run, in matches and in money play alike.
Hence a consequence that changes not the game but the way of learning it: reviewing your own games with a program has replaced the coach. What follows is about the programs themselves.
Where their strength comes from
A neural network, not a search. The strength of backgammon software lies in the neural networks’ weight tables, the result of extensive training. Without those tables the same program plays no better than a human novice.
Hence an unexpected property: a bot needs to know nothing about its opponent. Snowie and GNU play at a world-class level knowing nothing about whoever sits across the board.
Where the network is especially good is positional and strategic judgment. It is weaker in technical positions — bearing in against an anchor, for instance. The admission exists from the other side too: Mike Corbett’s problem collection is made of positions so hard that even Snowie gets them wrong.
XG, GNU and Snowie: how they differ
XG is the world benchmark. It is the absolute reference for analysis: it achieves a better ratio of playing strength to analysis speed than any other program available, and the best players check every game against it. A game with “zero errors” against XG is, by any human standard, essentially a perfect game.
GNU has overtaken Snowie. Snowie has long gone without an update, and GNU’s level has passed it. GNU’s four-ply assessment is usually better than Snowie 4.5’s. GNU is also harsher on grades: an Advanced grade in GNU could easily be an Expert or World Class grade in Snowie 4.
They also differ on specifics, which is worth knowing: in one position GNU prefers
6/3, 13/11 by a much smaller margin than Snowie, and Snowie has a positional
reason — the builders it leaves on the eight and seven points. Short Snowie
rollouts in another position show that White should not double after 24/23, 13/10, but should after either of the other plays.
How to use them
The order is plain:
- play as many games as possible, including against a strong program — the effect on your game is called very high;
- upload your matches into XG, Snowie or GNU afterwards — online platforms allow it;
- read the rollout equity, understand the reasoning behind the computer’s move, and compare it with your own;
- look at where exactly the equity was lost — the program shows it move by move.
A review after key sessions accelerates progress considerably, and it is available to anyone: software analysis gives self-improvement without an expensive tutor.
The mobile version. XG Mobile has strong neural-network AI, deep position analysis, and a tutor mode that points out mistakes outright and suggests better moves — sometimes moves a seasoned player would not have considered. It is the strongest option for serious learning, with one caveat: iOS remains the most stable option, and availability on Android has become inconsistent.
What a program does not give
One limitation matters more than the rest. An expert-level program gives the right answer as a number — but it will not say in words why that move is right. The numbers have to be read, and understanding does not follow from them automatically.
Hence the practical advice above: not “look at the evaluation” but read the equity, understand the reason, and compare it with your own.
The second limitation is about the human across the board: the human factor is knowing that neither we nor the opponent is XG. Play that is optimal by the program is not always optimal in practice against a weak opponent.
What it changed in the literature
Modern books written with XG and GNU analysis are more accurate than older texts, and that is not a courtesy: a program analyses any position more precisely than a book can, because a book gives worked examples and a program gives the position itself. Bill Robertie is named as one of the first players to integrate neural-network insights into his game.
Reviewing a game is on the strategy page, and the pip count these programs show in every position is on the pip count page.
Comments