An amateur player managed to beat an AI at the game of Go after identifying a flaw in the latter’s strategy. Proof that these systems are still limited to the training they have received.
By exploiting the error of the AI, Kellin Pelrine beat it in 14 out of 15 games.
A flaw spotted by… another program
Along with a team of researchers that he is part of, Pelrine has developed a computer program with the goal of looking for weaknesses in the AI strategy that a human player could take advantage of. For this, he made more than 1 million games against the AI called KataGo. The latter is based on the techniques used by DeepMind in the creation of AlphaGo Zero, the AI that beat the world’s best Go player in 2016 and caused him to announce his retirement three years later.
To beat her, Pelrine managed to surround groups of stones (players’ pieces) of the AI by drawing her attention to other areas of the goban every other move. The algorithm didn’t realize that its stones were surrounded, because it was focusing on another place on the goban (the Go game board). The researcher also applied this strategy to another AI specialist in the game of Go, Leela Zero, and won again.
AIs are still limited
According to the amateur player, this strategy would have been spotted very quickly by a human, because he would have realized the stones surrounding his own. This demonstrates that AIs can still have significant shortcomings and even blind spots, as is the case here. They are therefore not able to go beyond what they have been taught during their training.
Of course, these gaps can be filled through new training, but we observe the limits of these systems. Furthermore, algorithms can be prone to the same errors or biases as the humans who trained them.
Sources: Engadget, The Register
4