Bots, Neural Networks, and RNG Integrity

Backgammon was the first complex two-player game in which a computer demonstrably exceeded the best human players. The proof was published in 1992, four years before Deep Blue beat Kasparov and twenty-four years before AlphaGo, and it came from a single neural network trained by self-play at IBM Research. The lineage from that program — Gerald Tesauro's TD-Gammon — through Jellyfish, Snowie, GNU Backgammon, eXtreme Gammon (XG), BGBlitz, and the modern open-source engines, defines the technical history of computer backgammon.

This page traces that lineage, explains how modern bots are evaluated, documents the original GamesGrid bot family of 1996–2008, and sets out the platform's stance on random-number-generator integrity — including the Mersenne Twister algorithm we use and a documented contrast case from the operator history of online backgammon.

The companion page Performance Ratings (PR) & ELO covers the standard skill-measurement framework that the entire modern competitive ecosystem uses.

1. The Engine Timeline

Backgammon's bot history runs in five distinct phases. Each phase corresponds to a different fundamental approach to position evaluation.

1.1 Pre-neural: BKG (Hans Berliner, 1979)

The first computer program to defeat a reigning world champion at any complex board game was BKG 9.8, written by Hans Berliner at Carnegie Mellon. In 1979 it beat Luigi Villa, the reigning World Backgammon Champion, in a 7-point money match in Monte Carlo. The win was widely (and correctly) attributed in part to extraordinary dice — Berliner himself wrote that the program was probably weaker than Villa over a longer sample. But the result was undeniable: the era of "computers can't play backgammon" was over before the next phase even began.

BKG evaluated positions with hand-coded heuristics: features like prime length, blot exposure, and pip count, weighted by parameters tuned by Berliner against representative test positions. There was no learning component. It was the strongest non-learning program ever built for the game.

1.2 TD-Gammon (Tesauro, IBM Research, 1992)

The breakthrough came from a different paradigm. Gerald Tesauro at IBM Research applied temporal-difference reinforcement learning to backgammon. His program, TD-Gammon, was a feedforward neural network with a 198-unit input layer encoding the board (4 units per point per colour for checker-count distribution, plus 6 additional units encoding the bar, borne-off, and side-to-move) feeding into a single hidden layer of 40 to 80 sigmoidal units depending on the version (TD-Gammon 2.1 used 80 hidden units). It was trained by self-play — the program played itself for hundreds of thousands of games, and after each move adjusted its weights to make its evaluation of the position before the move closer to its evaluation of the position after the move.

The remarkable result, published in 1992 and refined through 1995, was that TD-Gammon discovered competitive openings that contradicted decades of human theory. Rollout analysis catalysed by TD-Gammon, and consolidated by Jellyfish and Snowie in subsequent years, overturned the 1970s slot-heavy consensus on several opening rolls (notably the 2-1, 4-1, and 5-1) in favour of split-builder plays. The standard played by every serious tournament player today for several opening rolls was set, in part, by this first wave of neural-net analysis. Cf. the opening rolls page.

1.3 Jellyfish, Snowie, GNU Backgammon (1994–early 2000s)

The first commercial NN-based program was Jellyfish, released by Fredrik Dahl in 1994. Jellyfish followed the TD-Gammon-style architecture (a feedforward network with a single hidden layer) and was the strongest publicly available program for several years. Snowie, released in 1998, used a more sophisticated network and added a polished cube-action analyser; it was the dominant commercial engine of the late 1990s.

GNU Backgammon (also called GNUbg or gnubg) is the open-source heir of that lineage. Initial development began in the late 1990s; major releases stabilised in the early 2000s. GNUbg's evaluation is a multi-layer feedforward network trained on rollouts of millions of positions; its 0-ply, 1-ply, 2-ply, and 3-ply evaluation modes give a range of speed/accuracy trade-offs that competitive players use for analysis. GNU Backgammon also publishes the standard reference Match Equity Table (Rockwell-Kazaross MET) used by the wider community.

1.4 eXtreme Gammon (XG) and Beyond (2009+)

eXtreme Gammon, released in 2009 by Xavier Dufaure de Citres, is the current world reference standard. XG combines a deeper neural network with optimised rollouts, an integrated cube-action analyser, and the most-used analytical user interface in the competitive scene. XG2 (the major revision) is the bot most often cited when modern PR ratings are reported. Tournament-grade Performance Ratings are routinely calibrated against XG2 rollouts at 4-ply truncation.

BGBlitz, by Frank Berger, is an independent NN engine in active development. Wildbg, an open-source Rust-based engine released in 2023, is notable less for raw playing strength than for training transparency — the training pipeline is publicly auditable, and the engine has become a research platform for evaluating different network architectures and training strategies. The transparency philosophy maps directly onto our own RNG and analytical-stack documentation, treated below.

2. How Bots Analyse Matches: The FIBS / GamesGrid → GNUbg Pipeline

The standard workflow in competitive backgammon — from 1996 through to today — is to play matches online, export them in a standardised text format, and analyse them in a neural-net engine after the fact. The original pipeline:

1. Play the match on FIBS, GamesGrid, or a similar server.
2. Export the match to a text file in SGF or JF (Jellyfish) format. FIBS supported export via the oldboard command; GamesGrid had its own export functionality.
3. Import the match into Jellyfish, Snowie, GNU Backgammon, or (later) XG.
4. Analyse at a chosen evaluation depth — typically 2-ply for daily review, full rollouts for serious study.
5. Report receives error rates per move, per cube decision, and an aggregate Performance Rating for the match.

The GamesGrid platform, in its 1996–2008 incarnation, was unusual among commercial servers in that match export was a first-class supported feature. The community of strong players who established their reputations on GamesGrid — playing thousands of rated matches and analysing every one against GNU Backgammon and Snowie — formed a substantial portion of the early-2000s competitive elite.

The 2026 platform continues this. Every match played on GamesGrid is exported on demand in standard formats and can be imported into GNU Backgammon, eXtreme Gammon, or BGBlitz for independent post-game review. There is no proprietary lock-in. The position-evaluation oracle is the player's own choice.

3. The Original GamesGrid Bot Family (1996–2008)

The 1996–2008 GamesGrid platform shipped a graded family of bots, all built on a fork of GNU Backgammon by GamesGrid Engineering (the engineering arm of CyberArts Inc., the parent company). Modifications to the engine — including improvements to the n-ply evaluation algorithms — were contributed back to the upstream GNU project, making GamesGrid one of the rare commercial operators that materially advanced the open-source backgammon ecosystem.

The bot graduation was achieved by inducing errors of varying magnitude rather than by training weaker networks from scratch. All bots used the same underlying GNUbg-derived neural network; the weaker bots simply made suboptimal plays at a controlled frequency. The original published ratings:

The bots played 2,000 to 4,000 matches per day across the platform, with maximum match length of 9 points. They played both standard backgammon and Nackgammon (the Nack Ballard opening-position variant). Match invitations were rate-limited by player rating: GG Otter accepted invitations only from players rated below 1800, GG Weasel only from players rated below 1600, and GG Chipmunk only from players rated below 1400.

MrHyperBot — a different architecture

Separately, MrHyperBot played the hypergammon variant — the three-checker speed variant where each player starts with only three checkers on the 24-, 23-, and 22-points. MrHyperBot did not use a neural network. Instead, it used an exhaustive position database containing the computed best play for every possible hypergammon position, developed by Hugh Sconyers (handle "hugh" on GamesGrid). Hypergammon's smaller state space made full game-theoretic solution computationally tractable, and Sconyers's database remains the canonical reference for the variant.

Xbot — Paul Magriel's bot

A separate independent bot, Xbot, was operated on GamesGrid by Paul Magriel (handle "X22"). Magriel — author of Backgammon (1976), the canonical positional theory text, and New York Times backgammon columnist 1977–1980 — ran a neural-network bot for small-stakes money matches up to 9 points. The presence of Magriel as an active operator on GamesGrid was emblematic of the platform's cultural place in the 1996–2008 era of competitive backgammon.

The full bot family — GG Forever, GG Raccoon, GG Otter, GG Weasel, GG Chipmunk, MrHyperBot, and the spirit of Xbot — returns in the 2026 platform, recreated from their documented playing fingerprints and joined by a new generation of named bot opponents. Further detail will be published closer to launch.

4. Random Number Generation: Why It Matters

A backgammon server is, at its core, a dice generator. The technical decision about how the server produces random rolls has enormous implications for competitive integrity — and the history of online backgammon includes documented cases of operators who got this wrong.

4.1 Mersenne Twister (MT19937)

The GamesGrid 2026 platform uses the Mersenne Twister pseudo-random number generator, specifically the MT19937 variant developed by Makoto Matsumoto and Takuji Nishimura in 1997. The algorithm has:

• A period of $2^{19937} - 1$ — astronomically larger than the number of dice rolls in any conceivable platform lifetime.
• A 623-dimensional equidistribution to 32-bit accuracy — meaning successive blocks of 623 outputs are uniformly distributed in a high-dimensional sense, not just in 1D.
• Speed: ~$10^7$ rolls per second on standard hardware, fast enough that the RNG is never a bottleneck.

MT19937 is the default RNG in GNU Backgammon, Python's random module, and a long list of scientific Monte Carlo simulators. Two technical caveats are worth stating explicitly:

1. It is not cryptographically secure. An attacker who observes 624 consecutive outputs can reconstruct the internal state and predict future outputs. For dice generation in an authenticated game server that does not expose its raw state, this is not a relevant concern, but it does rule out using MT19937 as the sole entropy source for security-critical purposes.
2. It has known statistical weaknesses on linearity tests. Because the algorithm is built on linear feedback over the two-element field $\mathbb{F}_2$, it predictably fails certain linear-complexity tests within the TestU01 BigCrush suite. It passes Diehard and the standard NIST randomness batteries, and the failures are well-characterised and orthogonal to the use case here (uniform integer rolls of 1–6); but a complete account of MT19937 should acknowledge them rather than describe the algorithm as universally test-passing.

For dice generation, the period and equidistribution properties decisively dominate any linearity concern. Cryptographic alternatives (e.g., AES-CTR-DRBG) are also available, and the platform's audit trail records the algorithm and seed source for every match.

4.2 Seed Audit Policy

Every match on GamesGrid is associated with a server-generated seed. The seed is:

1. Logged at match start to an append-only audit log.
2. Derived from a high-entropy source (hardware RNG combined with system entropy pool) at the time of match initialisation — not predictable from prior matches.
3. Recorded with the match record, allowing post-hoc reproduction of the dice sequence by an independent auditor.

In practical terms: every dice roll in a GamesGrid match can be independently verified to have been generated by MT19937 from a particular seed at a particular time. The dice are not "fixed" against any player, and there is no algorithmic mechanism for tilting outcomes for or against specific accounts.

4.3 The Contrast Case: SafeHarbor Games

The history of online backgammon includes operators who departed from neutral-RNG policy. The documented case is SafeHarbor Games, which at various points operated rooms that deliberately reduced the frequency of doubles to appease players who complained about losing to "lucky" rolls.

The intention was player retention. The effect was an exploitable RNG: any player who knew the doubles distribution was skewed had a measurable expected-value edge over players who did not. In a real-money context, this constitutes a violation of competitive integrity — and the broader online-backgammon scene reacted accordingly. The episode is referenced in player forums of the period and is part of the operator history that informs the GamesGrid platform's documented commitment to neutral, transparently algorithmic dice generation.

The position of the 2026 platform is unambiguous: uniform distribution, named algorithm, audit-logged seeds. The dice are the same for every player, in every room, at every score.

5. The 2026 Platform: What's Public, What's Coming

The new GamesGrid retains the architectural choices that distinguished the 1996–2008 server — closed-client play with server-side state, transparent match export to GNU Backgammon and XG, a graded family of bots descended from the original GG bot cast — and rebuilds them on modern engineering. The specific feature set, including the structure of the Career Mode bot leagues and the new bot roster, will be published closer to launch.

What is committed to publicly today:

• Uniform-distribution dice via MT19937, with audit-logged seeds.
• Match export in standard formats (SGF / JF) for independent analysis.
• No proprietary lock-in on position evaluation. Players choose their own analytical engine.
• The legacy GG bot family returns, recreated from documented playing fingerprints.
• A new generation of named bot opponents spans the skill ladder from beginner to world-championship calibre.

The full performance-rating framework that the bot family operates within is on the PR & ELO page.

See Also

• Performance Ratings (PR) & ELO — skill measurement methodology.
• History — the 1996–2008 GamesGrid era and the post-2020 reacquisition.
• Mathematics hub — match equity, take points, cube theory.
• Glossary — formal definitions of PR, ELO, neural network, Mersenne Twister, Hypergammon, Nackgammon.

Footnotes