About

Who builds Raccoon, and why

Raccoon is developed by Tekst og tal, an independent research and development company registered in Denmark (CVR 34101302).

We are a one-person operation working on applied machine learning, currently focused on a single question: how much of a classical game engine’s search can a modern neural network absorb into its weights?

Why backgammon

Backgammon is an unusual target. It is stochastic — 21 distinct dice rolls separate every decision from the next — which makes deep search expensive and rules out the tree-search tricks that carry chess engines. It has a large, irregular action space: OpenSpiel encodes 1,352 distinct checker moves. And it has a doubling cube, an equity-based side game that most published game-AI work ignores entirely.

It is also a game where the state of the art is genuinely old. GNU Backgammon, still the strongest freely available engine, was built around a small neural network in the TD-Gammon lineage, trained in the 1990s, with modern strength coming from full-width search layered on top. The methods that reshaped computer chess and Go after 2017 — deep residual networks, policy-value architectures, distillation at scale — have never been applied to backgammon in a sustained way.

That gap is the opportunity: a well-understood benchmark, a strong reference opponent to measure against honestly, and a community of serious players for whom the existing analysis tools are dated.

How we work

Everything is measured, and everything is published. Each experiment states one hypothesis, fixes one primary metric before the run, and reports every checkpoint against it with sample sizes and confidence intervals. The write-ups include the experiments that failed, the conclusions we later retracted, and the measurement bugs we found in our own evaluation code. That is deliberate — in a domain where the effect sizes are small and the noise is large, a result without its error bar is not a result.

The code is open. The engine, training pipelines, and evaluation harnesses are MIT-licensed and public. Game rules come from OpenSpiel; every line of the ML and search code is written from scratch.

Training runs on Google Cloud — preemptible GPU instances for network training, Cloud Storage for datasets and checkpoints.

Where this is going

The near-term goal is an engine that plays money-game backgammon at the level of the strongest existing bots while evaluating positions far more cheaply, exposed through a UCI-style protocol so that analysis tools and GUIs can embed it. Longer term, the interesting product is analysis: helping competitive players understand their errors, in a domain where the incumbent tooling is two decades old.

Today the project is pre-product, pre-revenue R&D. The engine handles checker play cubeless; cube handling is the next milestone.

Get in touch

Collaboration, questions, or backgammon data: contact us.