Chess AI | Matt G

Play Against the AI

You play White. Type a move in SAN notation and press Enter or Move. The AI plays Black.

♜

♞

♝

♛

♚

♝

♞

♜

♟

♙

♖

♘

♗

♕

♔

♗

♘

♖

Your turn (White).

Move log

No moves yet.

Description

Classical layer:The engine uses minimax with alpha-beta pruning for move search — the same approach as Connect 4 Bot, scaled to chess's larger branching factor (~35 vs ~7). Alpha-beta cuts branches that can't affect the final decision, making depth 4–6 search viable.

Neural layer: A ResNet-style policy/value network (ChessNet) trained in two stages. First, supervised pre-training on real human games from HuggingFace. Then MCTS self-play to improve beyond human patterns — the same architecture AlphaZero uses. The network outputs a policy over all 4096 possible moves and a value estimating win probability.

Why chess: Chess is the canonical benchmark for game AI — an engine that beats its own hand-tuned evaluation with a learned one is a concrete, measurable result.

Architecture

  Chess AI — Two-Layer Pipeline
  ══════════════════════════════════════════════════

  Layer 1: Classical Search
  ─────────────────────────
  Board state
       │
       ▼
  ┌────────────────────────────────────┐
  │  Minimax + Alpha-Beta Pruning      │
  │  - explores game tree to depth N   │
  │  - alpha-beta cuts ~50–70% of tree │
  └──────────────┬─────────────────────┘
                 │
                 ▼
  Best move returned

  ──────────────────────────────────────────────────

  Layer 2: Neural Network (ChessNet)
  ──────────────────────────────────
  Board → 13 input planes
  (6 white piece types + 6 black + 1 turn plane)
       │
       ▼
  ┌────────────────────────────────────┐
  │  Entry Conv → 5× ResBlock (64 ch)  │
  │  Residual connections, BatchNorm   │
  └──────────┬───────────┬─────────────┘
             │           │
             ▼           ▼
  ┌──────────────┐  ┌──────────────┐
  │ Policy Head  │  │  Value Head  │
  │ 4096 logits  │  │  tanh → ±1   │
  │ (all moves)  │  │  (win prob)  │
  └──────────────┘  └──────────────┘

  Training:
  1. Supervised pre-training on human games (HuggingFace dataset)
  2. MCTS self-play — policy/value network guides tree search,
     self-play results refine the network

Architecture mirrors AlphaZero: a single ResNet with two heads replaces both the move selector and the evaluation function.

Dev Notes

Tech Stack

Python, PyTorch for ChessNet, python-chess for move generation and board representation, torch.onnx for export.

From Connect 4 to Chess

Connect 4 Bot proved the minimax + tree search approach. Chess's branching factor is ~5× larger — alpha-beta pruning goes from useful to essential.

Why Not DQN

DQN outputs one Q-value per action from a single state — poorly suited to chess's 4096-move action space. The policy/value architecture handles the full move distribution directly, which is why AlphaZero uses it.

Download

chessnet.onnx

ChessNet exported to ONNX — portable and runtime-agnostic. Load with onnxruntime in Python or any ONNX-compatible engine.

ONNX~14 KB

↓ Download chessnet.onnx