Azul Engine Analysis board
New: a neural network brain

The engine taught a neural network, and the network beat it.

It learned from 1.5 million positions the engine played against itself, then improved by playing itself. It took 70% of the points in 3,000 games against the hand-tuned engine at equal thinking time.

See the results Neural network: 2,088 winsClassic: 879 wins

From the dev log

All posts
Machine learning

It taught itself: one generation of self-play

The network played 10,000 games against itself and a new network learned from them: +82 Elo over its parent. Then it levelled off.

 Machine learning

What the network looks at

Error by stage of the game and permutation importance on 105,000 positions it never trained on. It leans on the scored-now board and nearly ignores the tiles on the table.

 Performance engineering

Speed is strength: an NNUE-style evaluator in Go

Updating only what a move changes, in exact integer arithmetic, made the network three times faster and worth 50 Elo more, with the same weights.

Import / export

How to use the board

Play a move
Click tiles in a factory or the center, then click a highlighted pattern line or the floor.
Play the engine
Play vs engine gives you the side to move; the engine plays the other with the brain and thinking time you picked. Undo takes back your last move. Review game then grades every move against the engine's choice.
Copy a real game
Edit board, pick a color, then click factory slots, line cells, wall cells and floor slots. The board tells you if a position is impossible.
Choose a brain
Neural network judges positions with a machine-learning model trained on 1.5 million self-play positions; Classic uses hand-written rules. The network wins two games in three; see the benchmark.
Read the engine
The number is the expected lead in points for the player to move. Chips show what the move does right now: wall points, bonuses, floor penalty.
Explore a line
Click an engine move, then step with ◀ ▶ or the arrow keys. Play to here makes those moves.
End of a round
Score round, then Deal next round for a random deal, or edit in the real one.
Copy from a photo
Photo shows a picture of a real board above the editor while you click its tiles in. Read with AI fills the board in from the picture with your own AI key; always check its work.
Share
Share copies a link that opens this exact position.

Keys

← → step through a line · Esc cancel · Ctrl+Z undo · ? this help

Display

Read with AI uses your own Google Gemini or Anthropic key. It is saved only in this browser and sent, with the photo, only to that provider.