Don't Go There: Investigating the Move Locality of Human Go Players

Abstract

The rise of superhuman Go-playing programs like AlphaGo has raised questions about what distinguishes the computer play style from the human play style. Here, we investigate the claim that AlphaGo tends to play more globally than humans, jumping around the board in moments when humans might play locally. Although this claim has been generally accepted by Go players, a thorough empirical study is missing from the literature. Using data from self-play AlphaGo games and human games from the KGS Go Server, we measure the distance from each move to the previous move as an indicator of players’ global perspective. We find that, on average, humans play significantly closer than AlphaGo to the previous move. Furthermore, we discover evidence of a strong positive correlation between the average distance to the previous move and player skill level. These findings support the claim that AlphaGo plays more globally than humans and suggest that a global perspective is one aspect of a strong Go player. Finally, we show that humans play more optimally when the optimal move is closer to the previous move; meanwhile, mistakes can be primarily attributed to moves played too close to the previous move, rather than too far. We interpret this result as evidence of a human preference for moves in the local region, which may be due to cognitive biases or patterns of visual analysis and perception. Our findings are relevant to the design of better computer Go programs, the study of cognitive biases, and the collective improvement of human Go players.