Human beats top-ranking AI at game of Go (with help from another AI)
Human beats top-ranking AI at Go (with help from another AI)
An amateur Go player has definitively defeated a top-ranked AI system in the game of Go.
Kellin Pelrine, a player who is one level below the top amateur Go ranking, beat KataGo, one of the world's top Go AI systems 14-1. Yup, he won 14 out of 15 games.
Although he did this on his own without direct computer support, he did use tactics that were suggested by an AI program.
The program, designed by a Californian research firm called FAR AI, played over a million games against KataGo and successfully found a "bling spot" that humans could capitalise on.
For the uninitiated, Go is an ancient Chinese game played on a 19x19 grid where players alternately place black and white stones in an attempt to surround their opponent's stones and enclose the largest possible amount of space.
Because of the way the game is played, the Chinese call it "围棋" which literally means encirclement board game.
Despite its relatively simple-sounding rules, Go is extremely complex. Unlike chess, the vast number of combinations makes it impossible for any AI system to assess all potential future moves.
Using this weakness, Pelrine defeated KataGo by making decoy moves in other corners of the board while he slowly strung together a large loop of stones to encircle his opponent's stones.
According to Pelrine, this strategy would have been quite easy to spot, but KataGo did not notice this vulnerability at all, even when the encirclement was nearly complete.
This victory highlights a fundamental flaw in most AI systems in that they can only understand situations that they have been exposed to and often cannot generalise and see patterns in ways that humans might find easy.
Source: Financial Times