View
Best moves need the network.
Moves & analysis
No moves yet. Each move appears here; choose one to see the board as it was then.
Best moves share of the network's 32 searches
The bar beside the board is the network's reading at 32 searches. The numbered markers are each bot's own estimate from its latest move.
Up to 100 games. Player 1 moves first, which is an edge, so swap sides to compare bots fairly. Varied openings start each game with 4 random pawn moves.
How to move
Click a cell to move your pawn, or a groove between cells to place a wall; on a touch screen tap a groove once to preview a wall and again to place it. With the keyboard, focus the board, move the cursor with the arrow keys, press H or V to choose a horizontal or vertical wall, P to go back to pawn moves, and Enter to play. Page Up and Page Down step back and forward through the game, and Home and End jump to its start and its latest position.
Games
Score = wins + half the draws, per board size. Under about 20 games, differences are noise. Games with the same player on both sides, or with a seat changed part-way, are listed but not counted. The last 200 games are kept.
| Game | Player 1 | Player 2 | Result | Moves | Link |
|---|
About the bots
Beginner and Normal search ahead using shortest-path distances: how far each pawn is from its goal row, and how walls change that.
Strong and Max search ahead with a neural network, a convolutional one. It reads the position and gives every move a score, and the whole position a chance to win.
What is a convolutional network?
It looks at the board through a small window, 3 × 3 squares, that slides over every square, so a pattern it knows in one corner it knows anywhere. Each layer looks at the layer before it, so deeper layers see further: a wall several squares away, or a race between the two pawns. As the window does not depend on the size of the board, the same network can play several board sizes.