About Armagammon

Armagammon is built around Arma, a backgammon engine that started in late 2024 as a research project.

Backgammon is a hard game because every decision is made under uncertainty. The dice keep changing what can happen next, so strong play depends on weighing probabilities, risks, and competing plans rather than calculating a single forced line. Arma is an exploration of how neural networks and search can handle decisions like these.

Learning from the board

One of the questions behind Arma is how much of the strategic structure of backgammon a neural network can learn directly from the raw board.

To explore that question, Arma’s main network receives only the board position as input, with no handcrafted strategic features or higher-level interpretation. Whatever strategy it uses must emerge from training.

That is the experiment: how much can a network learn about backgammon on its own, and how far can that approach go?

More than the best move

Finding strong moves is only part of the problem. Another is whether a neural network can model human decisions, rather than only approximate ideal play.

A weaker player does not simply play like a stronger player with random mistakes added. The mistakes themselves have structure: some positions are difficult at one level and easy at another, and different players tend to misunderstand different kinds of decisions.

Arma is exploring whether those patterns can be learned, producing models that play like people of different strengths rather than engines with artificial errors added afterward. Peter, one of the opponents in Armagammon, is an early attempt at designing a network that plays in a more human way.

How much, and why

An engine can tell us that one move loses 0.079 equity compared with another. For analysis, that is extremely useful, but it answers only one question: how much. The harder question is why.

The long-term goal is to find out whether the strategic reasons behind an evaluation can be identified too: not just which move is better, but what is happening in the position that makes it better.

From Arma to Armagammon

Armagammon started as a simple interface for playing against Arma. Over time, it grew into a full application for playing, analyzing, and studying backgammon, and it is where the research meets you: the opponents, the hints, the analysis, and the tutor all come from Arma.

If you find a position that Arma misjudges, or one you would like it to explain, we would love to hear about it at hello (at) armagammon.com.