More than a computer opponent with a difficulty setting
Many backgammon bots are essentially one computer opponent with different strength settings. Make it weaker and the same engine is simply instructed to make more mistakes.
GamesGrid is built differently. There are 512 virtual players, each with a name, a face, a country, a voice, a personality and a measured playing strength of their own. Their differences go beyond the profile you see on screen. They play differently, handle the cube differently and communicate differently.
And you can actually talk to them. Every member can hold an open conversation with you, during a match or afterwards.
This page goes under the hood. Here is how their playing strength, mistakes, personalities and conversations are built.
512 players. 513 with Guru
The GamesGrid club is built around a large and varied cast:
-
512 characters, plus GG Guru. 513 in total.
-
120 countries. The biggest contingents are Lebanon and France with ten each, followed by Denmark with nine, and Turkey and Germany with eight.
-
Eight levels of play, from Beginner through Grandmaster.
-
A recorded voice for each of the 512, speaking across 58 languages.
-
A full profile for every member, including city, age, biography, playing style, archetype and cube temperament.
-
A career of their own. In Career Mode they enter events, build records, earn promotion, suffer relegation and progress through the same world as you.
You can meet all of them on our Our Players page.
You can actually talk to them
A large language model powers open conversation with every member, during a match and outside one.
Ask about their background, challenge an opinion, talk about backgammon or let the conversation wander somewhere completely different. It is not a menu of preset answers.
Their personality helps shape how they respond. A reserved member does not talk like an outgoing one. Some will have much more to say than others.
Once you have unlocked a member by beating them in Career Mode, you can visit their profile and continue the conversation away from the board.
Eight levels, each with a measured strength
Strength here is not simply a position on a difficulty slider.
GamesGrid uses PR, Performance Rating, to measure playing strength. PR measures the average equity a player gives away per decision, multiplied by 500.
Lower is better. Zero means no errors according to the analysis engine.
It is the same measurement framework used to assess human backgammon performance, which means a virtual player and a human player can be discussed on the same scale.
Every level has a target PR range, and every member has an individual target within that range.
For scale, a world-class human plays somewhere around PR 3 to 5 across a tournament. The Supremo tier sits above that on purpose. Guru is different again. Guru is the engine at full strength.
One important caveat applies to humans and virtual players alike: a single match PR can vary considerably.
A match full of straightforward decisions gives fewer opportunities to make mistakes. A difficult match can expose more weaknesses. A level therefore describes performance over many games, not a promise about the next match.
How do you make a bot play like a person?
Building a strong backgammon program was solved decades ago.
Building one that plays weaker in a believable way is a much more interesting problem.
The goal was not simply to make some members stronger and others weaker. We wanted their mistakes to appear in the kinds of positions where human players of similar strength also make mistakes.
That is where much of the work went.
The easy solution, and why we rejected it
One simple way to weaken a computer opponent is to make it choose a worse move every so often.
The problem is that this can produce an opponent that plays several excellent moves and then suddenly makes a blunder that feels completely out of character.
That may create a weaker opponent statistically, but it does not necessarily create a believable player.
We wanted the mistakes themselves to make sense.
We measured how people actually go wrong
To understand that, we measured human play.
Around 200,000 decisions made by approximately 100 real players were analysed, including GamesGrid test matches, spanning PR 1.7 to PR 15.9.
Each decision was compared with the engine's evaluation.
The question was not simply how often somebody made a mistake. We looked at where mistakes happened, how expensive they were, which alternative move was chosen and where players were usually accurate.
That data became the foundation for the playing model used across the cast.
How a mistake actually happens
People do not decide to make a blunder. They misread the position.
GamesGrid models that idea.
A member looks at the engine's evaluation of the available plays through a controlled amount of imperfect perception, then chooses the move that appears best from that imperfect view.
That means mistakes are more likely to appear when several moves are close.
When one move is clearly superior, even a weaker member is more likely to find it.
A member can also be more or less attentive on a particular decision. How often that happens depends on the playing level.
This allows a Casual player to occasionally make a significant mistake without forcing every move to be slightly inaccurate.
The mistakes follow rules
A weaker player does not have to play badly all the time.
An easy game can still be a good game. If a member faces mostly straightforward decisions, that member may produce a much better PR than its long-term target.
There is no error budget. A member is not forced to make a mistake because it has played several good moves in a row. Earlier versions of the model included compensation behaviour, but that mechanism was removed.
Races and bear-offs are usually cleaner. In the human data, races and bear-offs account for around 15% of moves but less than 1% of the total equity lost. The members reflect that pattern.
The size of mistakes changes with strength. A strong human around PR 2.3 selects the engine's preferred move roughly 85% of the time. A weaker player around PR 11 does so roughly 69% of the time. When either player misses the best move, more than half of those choices are the second-best move.
The model uses those patterns to create weaker play without turning weaker members into random players.
Mistakes we deliberately prevent
Human players can make almost any mistake imaginable, but some errors make a virtual opponent feel less convincing rather than more convincing.
For that reason, GamesGrid deliberately limits certain behaviours.
A member who is clearly ahead will not suddenly throw away a large advantage with an obviously avoidable blot.
Each playing level also has a ceiling on how much equity one individual checker-play decision can lose.
Those limits are intentional. The aim is not to reproduce every possible human mistake. It is to create opponents whose overall behaviour remains believable.
Each member also has individual sharpness within its level, so two Intermediate players are not simply identical opponents with different names and faces.
The cube is a second skill, and it is tuned separately
Checker play and cube handling are different skills.
A player can move checkers well and still give away significant equity through poor doubles, takes and passes. Another player may understand the cube better than their checker play would suggest.
GamesGrid treats those skills separately.
Each level has its own cube temperament. Beginners are generally more reluctant to double and more willing to take. Stronger levels become progressively more accurate.
For the three weakest levels, cube decisions use the same imperfect-perception approach used for checker play, fitted to the way weaker human players actually double, take and drop.
A member's level is based on the combined cost of checker-play and cube errors.
Playing the checkers like an Expert is not enough if the cube is consistently handled like a Beginner.
They resign, and they judge your resignations
A member who is losing badly can offer to resign a single game, a gammon, a backgammon or the entire match.
You decide whether to accept.
Resignation behaviour is calibrated too. Weaker members tend to resign later and less accurately than stronger ones.
And the decision matters on your side as well.
Accept an offer worth less than the position was worth if play continued, and the post-match analysis can record that as your mistake, including the equity cost.
Knowing when to resign, and when to accept a resignation, is part of playing good backgammon.
GG Guru is the exception
One player in the club has no added noise, no attentional drift and no weaker cube temperament.
GG Guru plays the engine's preferred move every time.
That does not make Guru unbeatable. Backgammon still has dice, and even the strongest play loses matches.
It simply means Guru is not deliberately weakened in the way the rest of the cast is.
The engine underneath every player
Underneath all 513 members is GNU Backgammon, a neural-network backgammon engine developed as free software over more than two decades.
The backgammon engine runs inside your browser. Position evaluation and move calculation do not need to be sent to a remote analysis engine or wait in an external analysis queue.
The conversational LLM is a separate system from the backgammon engine.
GNU Backgammon performs two different jobs within GamesGrid.
First, it evaluates the position during play. The member's individual playing model then determines what that character does with that evaluation.
Second, GNU Backgammon is used for post-match analysis, at a consistent strength for every player.
GNU Backgammon is licensed under the GNU GPL v3. GamesGrid includes the required licence information and source links in Settings.
The history of the engine, and the path from earlier backgammon programs to GNU Backgammon, is covered on our GNU Backgammon page.
The dice are not part of the difficulty level
Playing strength changes how a member handles a position. It does not change the dice.
No virtual member is given advance knowledge of future rolls.
A Beginner and GG Guru use the same dice source, regardless of playing level, score or whether they are winning or losing.
Every completed match also contains a record of the rolls and moves, so the match can be replayed afterwards.
When online multiplayer opens, dice will be issued by the server rather than by either player's device. The exact verification system will be documented once that implementation is complete.
The underlying principle is simple: the distribution is not changed to soften a bad run, manufacture a comeback or influence a result.
The playing model changes decisions, not dice.
The same judge for everyone
The opponent model and the post-match analyst are separate.
Whoever you played, from the weakest Beginner to GG Guru, the review afterwards uses the same settings: two-ply evaluation, no added noise and the same move filter.
A weaker member does not receive a more forgiving assessment. Guru does not receive a stricter one.
The same standard is also used when we measure and calibrate the club members.
Every match can be exported as a standard .mat file, allowing you to open it in compatible backgammon analysis software and review the match independently.
What the review shows is explained on our Backgammon Analysis page. What PR, rating and the playing levels mean is explained on our AI & Bots Performance page.
What we are not claiming
A technical page should be clear about the limits as well as the capabilities.
Their backgammon strength does not learn or adapt to you. A member's playing model remains the same in your hundredth match as in your first. The conversational AI is a separate part of the system.
Their mistakes are not random. The variation in the playing model is shaped to resemble human misjudgement rather than choosing arbitrary legal moves.
GG Guru is not unbeatable. Guru plays the engine's preferred moves. The dice still decide plenty.
A member's PR is a target. It is not a guarantee for one individual match.
Features still in development are described as such. Multiplayer, tournaments, leagues, 2 vs 2 and Tutor Mode will be identified as available when they are actually released.
