No Thanks Predictors

Zack Tillotson

A set of optimizable predictors (aka players) for the game No Thanks!

npm install
npm test

---

Strategy for Reinforcement Learner

• LM101 Link
  • http://www.learningmachines101.com/lm101-025-how-to-build-a-lunar-lander-autopilot-learning-machine/
• Strategy
  • 1. Define state vector
  • 1. Create game simulator
    • What are results of actions
    • Who are the opponents
    • Reenforcement signal
      • Special feature => small when doing well and large when doing poorly
      • eg
        • RS = (us - best competitor) / (total score taken + 1) * 50 + 50
        • 0-100 while playing
          • 50 is tied with everyone
          • 100 is losing by most possible
          • 0 is winning by most possible
        • 10x when game over and have lost
        • 1/10x when game over and have won
  • 1. Develop control law
    • Linear combination of features
    • Prob(Action|State) = sum( Feature_n(state) * W(n) )
      • Features
        • Table
          • Pot
          • Card ?
            • Redundant when player card value considered
          • Cards left
        • For each player
          • Your money
          • Your card total
          • Card value to you
  • 1. Develop learning rule
    • Adaptive gradient decent
    • W'(n) = W(n) * (Action - Prob(action)) * Reenforcement signal