Figuring out why AIs get flummoxed by some games

"With its Alpha series of game-playing AIs, Google's DeepMind group seemed to have found a way for its AIs to tackle any game, mastering games like chess and by repeatedly playing itself during training. But then some odd things happened as people started identifying Go positions that would lose against relative newcomers to the game but easily defeat a similar Go-playing AI."

"A recent paper published in Machine Learning describes an entire category of games where the method used to train AlphaGo and AlphaChess fails. The games in question can be remarkably simple, as exemplified by the one the researchers worked with: Nim, which involves two players taking turns removing matchsticks from a pyramid-shaped board until one is left without a legal move."

"These differ from something like chess, where each player has their own set of pieces; in impartial games, the two players share the same pieces and are bound by the same set of rules. Nim's importance stems from a theorem showing that any position in an impartial game can be analyzed mathematically."

DeepMind's Alpha series AIs excel at games like chess and Go through self-play training, but researchers discovered critical failure modes when these AIs encounter impartial games such as Nim. Impartial games differ from chess because both players share identical pieces and rules. Nim, a simple game involving removing matchsticks from pyramid-shaped rows, represents an entire category of games where AlphaGo and AlphaChess training methods fail. These games require intuiting underlying mathematical functions rather than learning patterns through repeated play. Identifying these AI blind spots helps improve training methods and prevents critical failures as AI systems become increasingly relied upon for complex problem-solving across various domains.

Read at Ars Technica