Raccoon Backgammon

A modern neural engine for backgammon

Backgammon’s strongest engines were built in the 1990s. Raccoon is a rebuild with the methods that have since transformed chess and Go — a deep residual network trained by large-scale distillation. It plays stronger backgammon than GNU Backgammon’s own network, it is open source, and it runs entirely on your machine. No account, no server, nothing to install.

The board below is the engine itself, not a video of it. Roll, and it plays you.

The board above needs JavaScript and WebAssembly. Everything else on this site works without them, and the engine’s code also runs in a terminal.

What you can do today

• Play a full cubeless money game against the engine — full-size board.
• Analyse any position: ranked candidate moves, cubeless equity, and the win/gammon/backgammon split — open the analyser.
• Bring your own positions with XGID import and export, the format XG and GNU Backgammon already speak.
• Read and run the code — the engine is MIT-licensed Python and PyTorch, though the trained weights ship separately. How it works.

Everything happens on your machine: the network’s weights are downloaded once and evaluated locally, so nothing you play is ever sent anywhere.

Why build it

Backgammon was one of machine learning’s early triumphs. TD-Gammon reached expert play in 1992, GNU Backgammon followed the same recipe as free software, and the commercial engines that beat them are closed and paid. Then the field moved on — residual networks, large-scale distillation, search guided by learned values — and almost none of it was ever brought back to backgammon.

Raccoon’s goal is to close that gap in the open: a modern, measurably stronger backgammon engine that any player, teacher or tool author can use, with the method published rather than locked up. Every step is measured against a named opponent on a metric fixed before the run, and published in full — including the experiments that failed — at research.raccoonbg.com.

Where it stands

The first milestone is done. On the BGSage money benchmark Raccoon scores PR 0.95 against 2.14 for GNU Backgammon’s own network at equal search depth, and in direct play it wins +0.047 ± 0.003 points per game over 18,000 games. The evidence, in two tables.

The honest caveat: this is checker play only, cubeless. The doubling cube — half of competitive backgammon — and a search layer on top of the network are the next two milestones. See the roadmap.