Figuring out why AIs get flummoxed by some games
2 days ago
- #AI
- #DeepMind
- #Game Theory
- Google’s DeepMind Alpha series AIs mastered games like chess and Go by self-play training.
- Some Go positions can defeat AI but lose to human newcomers, revealing AI blind spots.
- A recent paper identifies a category of games where AlphaGo’s training method fails, exemplified by Nim.
- Nim is a simple impartial game where players take turns removing matchsticks from a pyramid-shaped board.
- Impartial games, like Nim, share the same rules and pieces for both players, unlike chess.
- Nim is critical because any impartial game position can be represented by a Nim configuration.