Feb 20, 2023 14:00:00

A person who overwhelmed the strongest Go AI appeared, and it was talked about that human beings won by exploiting AI's weaknesses

With the evolution of artificial intelligence (AI), the number of cases in which AI wins against professional players in intelligent games such as chess and shogi has increased, but in January 2016

Google's 'AlphaGo' won Go's professional player Two months later, he won against the world 's strongest shogi player . AI reigned as an overwhelming winner in the world of Go, such as the announcement of the new version 'AlphaGo Zero' that can be remembered and strengthened . Under such circumstances, an amateur Go player defeated the strongest Go AI in 15 games and 14 wins without direct computer support, and it became a hot topic as 'Humanity's victory over AI'.

Adversarial Policies in Go - Game Viewer
https://goattack.far.ai/adversarial-policy-katago#contents

Subscribe to read | Financial Times
https://www.ft.com/content/175e5314-a7f7-4741-a786-273219f433a1

David beats Go-liath - by Gary Marcus
https://garymarcus.substack.com/p/david-beats-go-liath

Go AI 'AlphaGo' developed by DeepMind acquired by Google is the world's strongest Go player Lee Sedol, who won four crowns in world wars and achieved an overwhelming record of 1000 wins in total No. 5 In 2016, it was described as 'for the first time artificial intelligence has surpassed humans in Go' because it won the match 4-1. Although Lee won the fourth round, he commented, ``I just won against the AlphaGo bug.'' There is an existence that can not be done, 'he hints that AlphaGo's overwhelming strength is one of the reasons for his retirement.

The world champion who lost to Go AI 'AlphaGo' retired as a player saying 'I can't beat AI'-GIGAZINE

With AI posing a considerable threat to Go players, American player Kellin Perrin , who is one level below the top of the amateur rankings, exploited a previously unknown flaw in Go AI. and beat AlphaGo's level Go AI in 14 out of 15 games. The AI defect itself was identified by computer analysis, but it was said that it was done without direct computer support during the game. 'This win highlights a common weakness in most widely used AI systems today, such as OpenAI's ChatGPT,' said Adam Grieve, chief executive of FAR AI , the research firm that discovered the AI flaw. I am commenting.

A program designed by FAR AI has played over a million games against KataGo , one of the top Go systems, to find a 'blind spot' that human players can exploit. In the past, a study conducted at Cornell University in the United States also found a method that ``can win 99% against KataGo opponents'', but Cornell University's tactics are special using KataGo's reinforcement learning method On the other hand, the FAR AI tactic is relatively easy for intermediate level players to beat the AI, and in fact used the same tactic to beat Leela Zero, another top Go system. .

An AI learning method that can win 99% against the strongest Go AI ``KataGo'' opponent is devised, it is too special to be effective only for AI opponents-GIGAZINE

The tactic Perrin used in practice was to 'slowly encircle one of the opponent's positions by forming a large circle of stones, while distracting the AI by striking in the other corner of the board.' . Even if the enclosure is almost completed, the Go AI does not notice the current disadvantage, and Mr. Grieve said, ``The strategy that easily beats AlphaGo is not an effective way to play Go.'' Human Go players In addition to showing a crushing defeat in the game with, Mr. Perrin commented, ``Humans should be able to easily see that they are in a disadvantageous situation without being fooled by the strategy of distracting attention.''

Although the adversarial policy easily beats KataGo, it has not learned how to play Go effectively. My co-author @5kovt playing as white absolutely crushes the adversarial policy⚫: https://t.co/N6oFjWCg56 pic.twitter.com/wYolFeUhoR

— Adam Gleave (@ARGleave) November 2, 2022

According to Grieve, the exact cause of Go AI's disastrous defeat remains a matter of speculation, and a likely reason is that 'the tactics used by Perrin are rarely used in practice. Therefore, the AI system was not sufficiently trained.' “It is common in AI research to find flaws in AI systems when they are subjected to attacks that exploit the AI's identity,” said Grieve. We are seeing large AI systems deployed at scale with little validation.'

Stuart Russell, a professor of computer science at the University of California, Berkeley, said, ``The discovery of a vulnerability in one of the most advanced Go AIs is an indication that the deep learning system that supports today's most advanced AI. It suggests a fundamental flaw: the system can only understand specific situations experienced in the past, and cannot generalize experiences and current situations in a way that humans can easily understand.' 'We may have rushed too quickly to give machines superhuman levels of intelligence, and our victory over Go AI is yet another example of that,' he added. .

At goattack.far.ai , which FAR AI has published about this research, you can see the details of what kind of training and analysis was done, as well as the tactics to actually beat AI one by one.