Machine Learning with Matches and Confetti
summary
A physical reinforcement-learning demonstration uses six matchsticks and labeled pots of white and red marbles to represent actions: picking one or two sticks. Winning actions are rewarded by adding same-color marbles and losing actions penalized by removing them, producing a repeatable period-three winning pattern with occasional resets when pots empty; scaling the setup to tic-tac-toe increases the draw space to about 384 distinct outcomes. The device reproduces classic reinforcement-learning behavior and converges after many iterations of self-play.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:25Reinforcement learning
- 3:52Nim setup
- 5:14Human vs AI demo
- 7:17RL between automata
- 10:50Results after games
- 12:39Game complexity
- 13:42Automation limits
- 14:43Comparison with modern algorithms
- 15:24Historical origins
related talks
15mMachine learning with matches and confetti
48mTraining a PowerPoint AI to Play Tic Tac Toe
36mLearning about machine learning from kids
28mA Reinforcement Learning Framework in Scala 3
49mML: Teaching machine to learn from its own mistakes (...)
31mDesigning game AI with reinforcement learning
39mLet LLMs Wander: Engineering RL Environments
46mAI Takeover