NeuralGo
Two small neural networks, one for Black and one for White, play 9×9 Go against each other and learn from the results.
Self-play board (9×9)
Turn: Black
Move #: 0
Captures — B: 0 · W: 0
Black model, chance Black wins: —
White model, chance White wins: —
Status: Idle
Last completed game
Game #: 0
Winner: —
Score (B−W): —
Training
How it works. Each colour has its own network: 81 board cells plus a side-to-move input, one tanh hidden layer,
and a sigmoid output that estimates the chance its colour wins. To move, the player tries every legal move one step ahead and
picks the position its network rates highest, plus the capture bonus. With chance ε it plays a random move instead. There is no
tree search. When a game ends (two passes or 81 moves, area scoring, komi 6.5), each network is trained by gradient descent on
the positions it saw, with the target 1 for a win and 0 for a loss.
Learning rate (α)
0.10
Hidden units (resets networks)
Exploration (ε)
0.10
Move delay (ms)
160
Capture bonus
0.15
Added to a move's score for each stone it captures. Higher values make both players prefer captures.
Results
Total games: 0
Black wins: 0
White wins: 0
Black win rate: —
Winner predicted correctly: —
Avg confidence: —
Training steps: 0
Per-model stats
Black model
Predictions: 0
Correct: —
Avg confidence: —
Training steps: 0
White model
Predictions: 0
Correct: —
Avg confidence: —
Training steps: 0
Training stats
Orange: Black model's estimate for Black. Blue: White model's estimate for White.
One 9×9 map per hidden unit. Orange = positive weight, blue = negative; darker = larger. Inputs are +1 for the model's own stones and −1 for the opponent's. The caption shows the unit's bias, output weight, and side-to-move weight.
Raw weights
—
Black model
Bias: 0.000
White model
Bias: 0.000