Neural Network Architecture

Architecture Components

Input Encoding (4 channels)

Board state expanded to 4 channels: invalid positions, holes, pegs, and a learned board-type embedding that adapts to different game variants.

ResNet + Squeeze-Excitation

10 residual blocks with SE attention. Each block: Conv3x3 → BN → GELU → Conv3x3 → BN → SE → DropPath. Channel attention emphasizes important features.

Multi-Head Spatial Attention

4 transformer-style attention layers with 4 heads each. Enables global reasoning about distant board positions - crucial for planning long jump sequences.

Dual Output Heads

Policy: 1800 logits (15×15×8 directions). Value: Solvability prediction [0,1] as auxiliary task to improve learned representations.

Forward Pass

The network processes a board state through the following stages:

Input: Board State [15 × 15]

↓ expand to 4 channels

[15 × 15 × 4] → Conv3x3 + BN + GELU → [15 × 15 × 128]

↓ 10× residual blocks

[15 × 15 × 128] → ResBlock+SE ×10 → [15 × 15 × 128]

↓ 4× attention layers

[15 × 15 × 128] → MultiHeadAttn ×4 → [15 × 15 × 128]

↓ split to heads

Policy Head

Conv1x1 → [15×15×64]

Conv1x1 → [15×15×8]

Output: 1800 logits

Value Head

GlobalAvgPool → [128]

FC → [64] → [1]