Go

Goal of the game

Two players, Black and White, take turns placing a stone on a free intersection of the board. The aim is to control the largest area: your living stones and the empty intersections that you alone surround.

How it works

• Black starts. A stone that has been placed never moves again, unless it is captured.
• The liberties of a stone are the empty intersections touching it (up, down, left, right). Connected stones of the same colour form a group that shares its liberties.
• A group deprived of its last liberty is captured: its stones are removed and become prisoners.
• Suicide is forbidden: you cannot play where your group would have no liberties, unless that move captures.
• Ko: it is forbidden to immediately retake a single stone that has itself just captured a single stone. Play elsewhere first (a “ko threat”), then retake.
• You may pass. Two consecutive passes end the game. You may also resign.

Chinese area scoring

Each side counts its living stones plus the empty intersections surrounded only by its stones. White receives a komi of 7.5 points to offset the advantage of moving first; the half point rules out any tie.

At the end, the AI estimates the dead stones (greyed out) and each side’s territory. Tap a group to change its status, then confirm. In case of disagreement, resume the game: under area scoring, playing inside your own territory costs nothing.

Controls

• P key: pass.
• “Estimate” button: shows the territory estimate during the game.
• On 13 × 13 and 19 × 19 boards on touch screens, tap once to aim, a second time to place the stone.

About the AI

Monte Carlo tree search (UCT) enriched with RAVE, which shares move statistics between simulations. Each node receives prior knowledge: captures, rescues, ladder reading, 3 × 3 patterns and edge lines. The simulated games are guided: defending and attacking groups in atari, local patterns, rejecting suicide, self-atari and filling in its own eyes. The Master level thinks for about five seconds and simulates tens of thousands of games per move on 9 × 9; its moves are rougher on large boards. The AI passes when no move is useful any more and resigns a hopelessly lost game.