🎲 BANK AI Trainer

Round 1 of 10 | Roll #0 | Phase: Safe Rolls

💡 Agent Learning: The AI uses Q-learning to improve its strategy. It learns when to bank based on bank total, score difference, round number, and roll count. More training episodes = smarter AI!

What is this?

This is an AI training tool for the dice game BANK. You can train a Q-learning agent to play the game intelligently, then test your skills against it!

How does it work?

• Q-Learning: The AI learns by playing thousands of games and remembering which decisions lead to winning.
• State Space: The AI considers bank total, score difference, round number, and roll count when deciding to bank.
• Exploration: Initially, the AI tries random moves. Over time, it uses what it learned.
• Reward Shaping: The AI gets rewarded for winning, banking at good times, and penalized for busting.

Game Rules (BANK)

• First 3 rolls (Safe Phase):
  • Rolling a 7: +70 bonus points
  • Rolling doubles: Add face value to bank
• After 3 rolls (Risky Phase):
  • Rolling a 7: BUST! Lose all unbanked points
  • Rolling doubles: DOUBLE the entire bank total
• Banking: Secure your bank total to your score (can only bank once per round)
• Winning: Highest score after all rounds wins!

Training Parameters Explained

• Learning Rate (Alpha):
  • Controls how quickly the AI updates its knowledge
  • Range: 0.01 to 1.0 (default: 0.1)
  • Higher (0.3-0.5): Faster learning, but may be unstable or forget old lessons
  • Lower (0.05-0.1): Slower learning, but more stable and reliable
  • 💡 Tip: Start with 0.1, increase if learning is too slow
• Exploration Rate (Epsilon):
  • Controls how often the AI tries random moves vs. using what it learned
  • Range: 0 to 1 (default: 0.3)
  • Higher (0.5-0.8): More random exploration, tries new strategies
  • Lower (0.1-0.3): More exploitation, uses learned strategies
  • Automatically decays during training (starts high, ends low)
  • 💡 Tip: Start with 0.3-0.5, let it decay naturally
• Opponent Type:
  • Random Agent: Banks randomly (30% chance each roll). Easy to beat, good for early learning.
  • Threshold Agent: Banks when total exceeds 150 points. Medium difficulty, teaches timing.
  • Expected Value Agent: Banks based on mathematical calculations. Harder opponent.
  • Mixed Opponents: Rotates through all types. Best for robust training!
  • 💡 Tip: Use "Mixed" for well-rounded AI that handles different strategies

Tips for Best Results

• Start with 1000 episodes to see basic learning
• Train for 5000-10000 episodes for competitive AI
• Train against mixed opponents for best results
• Save your AI at different training stages to compare!
• Watch the win rate increase as the AI learns!
• If win rate plateaus, try adjusting learning rate or training more episodes