AI Training Fails in Impartial Games

A recent study published in *Machine Learning* by Bei Zhou and Soren Riis indicates that the self-play training methods used for advanced AI systems like Google DeepMind's AlphaGo and AlphaZero are ineffective for a category of games known as impartial games. This limitation was demonstrated using the game of Nim, where AI performance significantly degraded as game complexity increased. For a five-row Nim board, the AI showed improvement over 500 training iterations, but for a seven-row board, performance gains ceased after 500 iterations, becoming indistinguishable from random move selection.

Impartial games, unlike games such as chess, feature shared pieces and identical rules for both players. Nim is a critical example, as any position in an impartial game can be represented by a Nim configuration. Optimal moves in impartial games can be determined by a parity function, which indicates the winning player. The AlphaGo approach, which relies on associating board configurations with win probabilities through self-play and random sampling, struggles to develop this functional understanding.

The research highlights a fundamental failure mode in current AI training paradigms when applied to environments where optimal strategies are derived from underlying mathematical functions rather than complex, probabilistic evaluations. This suggests a need for alternative training methodologies to address these specific types of game structures, impacting the broader application of AI in problem-solving.