← Backgammon Battles

How backgammon bots learned the game

Backgammon bots are programs that judge a position with a neural network trained mostly by playing against themselves. Since IBM's TD-Gammon in the early 1990s they have played at or near the level of the best humans, and today they matter less as opponents than as referees: they grade every move a person makes, which is what a performance rating is built on.

That arc, from curiosity to opponent to referee, is the reason this page exists on a site about playing. If you have ever looked at a computer review of your own game and wondered why anyone should trust the machine's verdict, the answer is in how these programs were built, and in what they still get wrong.

Find out exactly how good you are

Every roll is committed in advance and provable afterwards. Every match is rated. Every game can be replayed move by move with a computer's assessment beside your own. Nothing about how you are judged here is hidden from you.

Play ranked, free

Before the neural networks: BKG 9.8

The first famous backgammon program was BKG 9.8, written by Hans Berliner at Carnegie Mellon University. Berliner was a former world correspondence chess champion, and he noticed something that chess programmers of the time did not have to deal with: in backgammon you cannot search very far ahead, because every turn branches into every possible roll of the dice. His program leaned instead on positional judgment, a hand-built way of weighing features of the board such as blocking, safety and the race. In his own account of it, the program did well more by judgment than by brute calculation.

In 1979 BKG 9.8 played the reigning world champion, Luigi Villa, in an exhibition match in Monte Carlo, and won. Berliner described it as the first time a reigning champion of a recognised intellectual game had lost to a machine. It was also a short match, and anyone who plays backgammon knows how much a short match owes to the dice. The honest reading is that BKG 9.8 proved a program could play backgammon well enough to win against the best, not that it was the stronger player. That distinction between one result and a measured level of play is exactly the one the game keeps forcing on people, and the reason the skill or luck question needs more than one match to answer.

TD-Gammon: the program that taught itself

The breakthrough came from Gerald Tesauro at IBM in the early 1990s. TD-Gammon was a neural network that learned by temporal difference learning: it played game after game against itself, and after every move it nudged its estimate of the previous position toward its estimate of the new one. Nobody told it which moves were good. It started knowing only the rules, and the results of its own games slowly taught it what winning positions look like. Later versions added some hand-designed board features and a short look-ahead, and the strongest of them played close to the level of the world's top players.

Backgammon turned out to be unusually well suited to this kind of self-teaching, and the dice are the reason. A program playing chess against itself tends to fall into the same few lines over and over. In backgammon, the dice keep throwing the program into positions it would never have chosen, so it explores the whole game whether it wants to or not. The same randomness that makes a single game feel unfair to a human made the game an ideal classroom for a machine.

TD-Gammon became one of the most cited results in machine learning, and the self-play idea runs in a straight line to the programs that later mastered chess and Go. For backgammon players, though, its biggest effect was more immediate: it disagreed with the experts, and the experts eventually admitted it was right.

The opening moves a computer overturned

Cream's opening 4-1 played 24/23 13/9, the split TD-Gammon preferred to slotting the 5-point
The modern 4-1: 24/23 13/9. Before TD-Gammon, most experts slotted with 13/9 6/5 instead.

Before TD-Gammon, the standard expert play with an opening 2-1, 4-1 or 5-1 was to slot: bring a checker down from the 6-point to the 5-point and dare the opponent to hit it. TD-Gammon preferred the quieter split, moving a back checker from the 24-point to the 23-point instead. Tournament players tried the computer's move, found that it worked, and within a few years slotting with those rolls had largely disappeared from serious play.

The story did not end with a clean verdict, which is the most useful thing about it. For 4-1 and 5-1 the split is still the modern standard, as our pages on the 4-1 opening and the 5-1 opening explain. For 2-1, later and stronger analysis brought the slot back into the conversation, and the two plays are now close enough that the match score can decide between them. The 2-1 page sets out both sides. A bot changed opening theory, and then better bots refined the change. That is what it looks like when the evidence is real.

The engines that followed

Through the 1990s and 2000s a series of programs built on the same neural network approach and turned it into tools that ordinary players could own. Jellyfish, written by Fredrik Dahl, arrived in the mid-1990s and was the first strong program many club players ever used. Snowie followed later in the decade, and its error-rate style of grading a player's whole match is still the basis of how performance is measured today, including the PR number on this site. GNU Backgammon brought a strong engine to everyone for free as open-source software. eXtreme Gammon, which followed later still, is the program many serious players treat as the reference today.

The details of each engine differ: how deep they look ahead, how they handle the cube, how they count which moves go into an average. What they share is the method. A network trained on enormous numbers of self-played games estimates how often each side wins, wins a gammon or wins a backgammon from any position, and every judgment the program makes follows from those estimates.

What a rollout is, and why it is not the last word

When an engine's quick judgment is not trusted, it does a rollout: it plays the same position out to the end a very large number of times with different dice, using its own judgment for every move along the way, and averages the results. Clever tricks reduce the luck in those dice so that fewer games are needed for a reliable answer. A rollout is the closest thing backgammon has to a measured fact about a position, and it is where claims like "the split is better" actually come from.

It is still the bot's opinion of itself, played out many times. If the engine misjudges a type of position, the rollout inherits the misjudgment. In clear positions that hardly matters. In close ones, two engines or two rollout settings can disagree, and an honest analysis says so rather than declaring a winner. You will see that honesty throughout our opening pages, which separate the openings where the answer is agreed from the ones still argued.

From opponent to referee

Playing against a bot is good practice, and you can do that here with the practice bot, which never affects your rating. But the bigger change bots made is not that you can lose to one. It is that every decision you make against a person can now be checked against the strongest play available, so your skill can be measured directly rather than inferred from results that the dice distort.

On Backgammon Battles every completed match against a person is analysed by GNU Backgammon, and the review marks each of your mistakes and turns them into a single error rate. That is a very different thing from a win record. A win record tells you how the dice and your opponents treated you. An error rate tells you how well you actually played, and it means something after far fewer games, because it judges every move rather than only the final result.

The bot and the dice are separate things

One worry comes up whenever players hear that a computer is involved: does the machine have anything to do with the rolls? It does not, and it should not be able to. A bot chooses moves. The dice come from a separate source, and here every roll is committed before it is used and can be checked afterwards, whether your opponent is a person or the practice bot. How the dice are proved fair explains the mechanism, and are online dice rigged deals with why the game so often feels otherwise.

The same strength that makes bots good referees also makes them a temptation: a player who consults an engine during a rated game is no longer the one playing. That is one more reason a rating is only worth something on a site that takes the integrity of the whole game seriously, the dice included.

See what the referee thinks of your game

Play a rated game against a person and the engine reviews every decision afterwards, with your error rate at the end. Start as a guest; nothing to install.

Play a rated game free

Frequently asked questions

What was the first backgammon program to beat a world champion?

BKG 9.8, written by Hans Berliner at Carnegie Mellon University, beat the reigning world champion Luigi Villa in an exhibition match in Monte Carlo in 1979. It was a short match, and short backgammon matches carry a lot of luck, so the result was a milestone more than a proof that the program was stronger.

What is TD-Gammon?

TD-Gammon is a backgammon program built by Gerald Tesauro at IBM in the early 1990s. It used a neural network trained by temporal difference learning, playing games against itself and adjusting its evaluation after every move. It reached a level close to the best human players and changed how experts play several opening rolls.

Are backgammon bots better than humans?

The modern engines are generally regarded as stronger than any human at checker play and cube decisions, which is why serious players use them to analyse their games. They are not perfect: in close positions, different engines and different rollout settings can still disagree, and those disagreements are reported honestly by good analysis tools.

Can a backgammon bot control the dice?

A bot only chooses moves. On Backgammon Battles the dice come from the same server-side source whether you play a person or the practice bot, each roll is committed before it is used and can be verified afterwards, and the bot cannot see future rolls.

© 2026 Backgammon Battles