2048

SCORE

Seed:

Use ↑ ↓ ← → or W A S D to move Swipe to play!

🤖 AI Player

AI Model:

Speed: 5 moves/sec

Loading AI model...

📚 About the AI Bots

🎲 Random (Baseline)

Picks moves randomly from available legal directions. Used as a baseline to compare other strategies.

Average Score: ~1,000-2,000
Max Tile: Usually 128-256
Speed: Instant

⚡ Expectimax (Rust) - RECOMMENDED

Bitboard expectimax with a tuned heuristic. Uses pre-computed lookup tables for O(1) move operations and searches the game tree to maximize expected value.

Heuristic: Empty cells, available merges, monotonicity & tile-sum penalties
Search: Expectimax with depth adapting to board complexity
Result: Reliably reaches the 2048 tile (often 4096+)

🎲 MCTS (Monte Carlo)

Monte Carlo Tree Search — runs random rollouts from each possible move and picks the one with the highest average outcome.

Strategy: 200 random simulations per move
Max Tile: Typically reaches 1024-2048
Advantage: No training required, works immediately

🎯 DQN Shaped

Dueling Deep Q-Network trained with advanced reward shaping (corner bonus, monotonicity, empty cell incentives).

Training: 100,000 episodes via reinforcement learning
Architecture: Dueling DQN (512→512 shared, value/advantage streams)
Framework: PyTorch on Apple M4 GPU

🖼️ CNN DQN

Dueling CNN that learns spatial patterns on the 4×4 board with one-hot encoded tile channels.

Training: 100,000 episodes with reward shaping
Architecture: 3 conv layers (128ch + BatchNorm) + Dueling head
Advantage: Learns position-aware features automatically

🔧 Implementation Details

Expectimax Rust: Uses bitboard representation (u64 with 4 bits per tile) and pre-computed lookup tables (65,536 entries) for instant move calculations — including a per-row heuristic table scoring empty cells, merges, monotonicity and tile sums. Search depth adapts to board complexity.

Risk-Averse Search: At chance nodes, blends average and minimum: (1-α)×avg + α×min where α=0.25, making moves more conservative.