AI 2048 教练：通过屏幕镜像实时指导 iPhone 游戏的神经网络系统

This is an innovative AI game coach system for the iPhone version of 2048, developed by Illya97 and released on GitHub on July 12, 2026 (link: https://github.com/Illya97/2048). It captures the iPhone screen via AirPlay mirroring, recognizes the game board in real time, and uses a self-learned n-tuple neural network combined with expectimax search to provide optimal sliding suggestions within 45 milliseconds, displayed as on-screen arrows.

Key technologies involved: 2048 game screen recognition, n-tuple neural network, TD/TC reinforcement learning, expectimax search, real-time system integration.

The AI coach system assists iPhone 2048 players with a 4-step workflow:

1. Capture screen via AirPlay mirroring.
2. Recognize 16-grid game state.
3. Compute optimal move using trained n-tuple neural network + 4-step expectimax search.
4. Show suggested direction via green arrows.

The core decision component is an n-tuple neural network trained on 9 million self-play games using TD/TC reinforcement learning, accelerated by numba with a pure Python alternative available.

Real-time Screen Recognition

• Adaptive palette: Learns any game theme's color config.
• Anti-interference: Robust to mirroring noise and fast-slide blurriness.
• Full grid: Accurately identifies all 16 cells.

N-tuple Neural Network

• Architecture: 9 templates ×8 symmetries, numba-accelerated.
• Training: 9M self-play games with TD/TC learning.
• Search: 4-step expectimax (≈45ms per move).

Smart State Management

• Round prediction: Shows estimated remaining rounds and补全 missing steps.
• Restart advice: Warns of weak rounds at ~300 steps if below average.
• End control: Suggests 1024+1024 merge only if expected score >47k; manual end via 🏁 button.
• Endgame protection: Hides outdated arrows when multiple high-value blocks exist.

The system uses intuitive cues:

• Green arrows: Recommended sliding direction.
• No arrows/⏳ Thinking...: System computing move.
• Optional sound prompts: Notify when suggestion is ready.

These help players quickly understand system state and follow recommendations.

Environment Requirements

• Python ≥3.10; iPhone + computer on same Wi-Fi.

Installation

1. Clone repo: git clone https://github.com/Illya97/2048.git && cd 2048
2. Install dependencies: pip install -r requirements.txt

AirPlay Setup

1. Install receiver (e.g., LetsView) on computer.
2. iPhone: Control Center → Screen Mirroring → Select computer.
3. Open 2048 game.

Calibration & Launch

• Calibrate: python -m coach --calibrate (verify 16-grid recognition).
• Start: python -m coach.

Custom Theme Support

1. Capture board screenshot as shot.png.
2. Learn colors: python -m coach --learn shot.png (input grid numbers to train config).

Pre-trained Weights

• Stored in weights/ntuple.npz (not in repo; expectimax as backup).

Continue Training

• Command: python scripts/train_ntuple.py --tc --resume --games 1000000
• Options: --parallel (all CPU cores), --warm-from path.npz (hot-start).

Benchmark

• Command: python scripts/benchmark.py --farm --agent ntuple --nt-depth4 --games100 --seed200

Evaluates move quality and speed.

Significance

• Integrates CV, RL, and real-time systems into a practical tool.
• Applications: Game AI research, real-time image processing, RL demonstration, cross-device interaction.
• For players: Learning tool for high-score strategies.

Limitations

• Single monitor required.
• Re-calibrate if mirror window is moved.
• Classic theme: Blocks >4096 share same color.