In 2016, a computer named AlphaGo made headlines when it beat then-world champion Lee Sedol at the popular ancient strategy game Go. A “superhuman” artificial intelligence developed by Google DeepMind has beaten Sedor in only one of his five rounds and Garry Kasparov’s 1997 chess defeat against IBM’s Deep Blue. is compared with Go, in which players face off with the goal of occupying space on the game board by moving black and white pieces called stones, was seen as a more difficult challenge for mechanical opponents than chess.
Many woes about AI’s threat to human ingenuity and livelihoods followed AlphaGo’s victory, much like what is happening now with ChatGPT and its ilk. However, at a press conference in 2016 after the loss, a subdued Sedol offered positive comments. He said. “AlphaGo made me realize that I need to study Go more.”
European Go champion Huang Hui had lost five private rounds to AlphaGo a few months earlier. Wired The match made him see the game as “completely different”. Wired.
It can be difficult to formally track the tedious process of human decision-making. But decades of recording the movements of a professional Go player have given researchers a way to assess the human’s strategic response to his AI’s taunts. Fan Hui’s improvement after facing the AlphaGo challenge was no mere fluke, new research confirms. We now have access to data detailing the moves made by , developing new strategies that lead to higher quality decisions in gameplay in a very human way. Confirmation of changes in human gameplay comes in findings published on March 13th. Proceedings of the National Academy of Sciences.
David Silver, principal research scientist at DeepMind and leader of the AlphaGo project, said, “It’s amazing how quickly human players have adapted to incorporate these new discoveries into their play. It is.” “These results suggest that humans adapt and build on these discoveries, greatly increasing their potential.”
To determine whether the advent of superhuman AI has driven humans to generate new strategies for gameplay, Minkyu Shin, assistant professor of marketing at the City University of Hong Kong, and his colleagues used 580 I used a database of 10,000 hands. From 1950 to 2021. The record, maintained on the website Games of Go on Download, reflects every move in the game of Go played in tournaments dating back to the 19th century. Researchers began analyzing the game because 1950 was the year the rules of Go were established.
To comb through the massive record of 5.8 million game moves, the team first created a method to assess the quality of each move’s decision making. To develop this metric, the researcher used yet another of his AI systems, his KataGo, to compare the win rate of each human decision to his AI’s decision win rate. This large-scale analysis included his 10,000 simulations where the game unfolds after each of 5.8 million human decisions.
After performing a quality assessment for each human decision, the researchers found a way to pinpoint exactly when a human decision in a game was new, i.e. not previously recorded in the history of the game. developed. Chess players have long used a similar approach to determine when new strategies emerge in game play.
In a Go game novelty analysis, researchers mapped up to 60 moves per game and marked when new moves were introduced. For example, if in one game he appears in move 9, and in another game he does not appear until move 15, the former game will have a higher Novelty His Index His score than the latter. Singh and his colleagues found that since 2017, most moves the team defined as novel occurred by his 35th move.
The researchers next asked whether the timing of novel moves in gameplay tracked improvements in the quality of decision-making, i.e., making such moves actually increased a player’s advantage on the board and their chances of winning. In particular, they examined whether AlphaGo’s decision quality improved after beating Sedol, a human challenger in 2016, and another set of human challengers in 2017. wanted to know what happened.
The team found that the quality level of human decision-making remained fairly constant for 66 years before an AI beat a human Go champion. After that fateful 2016-his 2017 period, decision quality scores began to rise. Humans were making better game play choices. Probably not good enough to beat superhuman AI consistently, but good nonetheless.
Novelty scores also skyrocketed in 2016-2017 as humans introduced new moves into the game early in the game play sequences. Evaluating the association between new moves and better decision-making, Shin et al. did not contribute to high decision-making. After these groundbreaking AI wins, the novel moves that humans introduced into the game, on average, contributed better decision quality scores than the already known moves.
One possible explanation for these improvements is that humans memorized movements for new play sequences. In this study, Singh and his colleagues also assessed how much memory could explain the quality of decision-making. Researchers find memorization does not fully explain the improvement in decision-making quality and is ‘unlikely’ to underlie the increase in novelty seen since 2016-2017 Did.
Murat Kantarcioglu, a professor of computer science at the University of Texas at Dallas, puts these findings together with the work he and others have done: “Clearly, AI can improve human decision-making. “It helps to do that.” Kantarcioglu, who was not involved in the current research, said that AI’s ability to handle a “vast search space” such as all possible moves in a complex game such as Go could lead to AI “providing new solutions to problems.” “We can find an approach.” For example, AI that flags medical images as suggestive of cancer could lead clinicians to look more closely. “This will help the person become a better doctor and prevent such mistakes in the future,” he says.
Kantarcioglu adds that the problem the world is currently seeing with ChatGPT is that of making AI more trustworthy. “I think that’s the main challenge,” he says.
In this new phase of ChatGPT and other AI concerns, the findings offer a “hopeful perspective” on the potential for AI to become an ally rather than a “potential adversary in our journey towards progress and improvement.” ‘, said Shin and his co-authors.wrote by email to Scientific American.
“My co-authors and I are currently conducting online lab experiments to explore how humans can improve prompts and achieve better results from these programs,” Singh says. “Instead of seeing AI as a threat to human intelligence, we need to embrace it as a valuable tool that can enhance our capabilities.”
