Checkers AI: Alpha-Beta vs. Reinforcement Learning
Checkers AI ♟️
A Classic Board Game But With AI Players
This project implements an intelligent checkers engine where you can:
- Play against another human
- Play against an Alpha-Beta pruning AI
- Train and play against a Reinforcement Learning agent
- Simulate AI vs. AI matches
It runs fully in the terminal using ASCII graphics, and strictly follows official checkers rules.
🧠 AI Techniques Implemented
Alpha-Beta Player
- Implements depth-limited minimax with alpha-beta pruning.
- Evaluation function values:
- Normal pieces = 1 point
- King pieces = 3 points
- Offers consistent, rule-based performance.
- Depth is configurable (default: 4); higher depths slow down execution.
Reinforcement Learning Player (Q-Learning)
- Learns optimal moves through self-play using Q-learning.
- Reward structure:
- +1 for win
- -1 for loss
- 0 otherwise
- ε-greedy policy for balancing exploration and exploitation.
- Stores training data in a persistent Q-table using
pickle.
🎮 Game Mechanics
- 8x8 board with diagonal movement, forced captures, and king promotion.
- Enforces all official rules including multi-jumps.
- Promotes pawns to kings on the far row, allowing backward movement.
- Auto-executes forced single capture moves for streamlined play.
📊 Experimental Results
- Initial results showed ~50/50 win rates between the agents.
- After extended training (e.g., 10,000 episodes), the RL agent outperformed the alpha-beta agent in most matches.
- Deeper alpha-beta search (e.g., depth 10) became computationally expensive and still fell short strategically.
- RL agent proved more adaptable and learned long-term strategies not captured by the evaluation function.
🔍 Key Takeaways
- Alpha-beta is fast and consistent, but limited by its static evaluation function.
- Q-learning, while slower to train, adapts better over time and can outperform rule-based agents.
- Training and simulation features are all built-in and can be triggered via the CLI game menu.
🛠️ Tech Stack
- Python: Game logic, CLI interaction, AI implementation
- Q-Learning: Reinforcement learning algorithm
- Pickle: Persistent Q-table storage
- ASCII UI: Visual board representation in terminal