We obviously know Deep Blue and AlphaZero, two artificial intelligences that have defeated high-level chess players and have demonstrated their ability to master extremely complex games.
But scientists at the University of Toronto have found that these AIs adopt a “alien style of play”, not very human, if you will, and which ultimately does not serve to improve the teaching of chess to human players.
For AI to be more effective and to be able to truly help a player to improve, they would still have to be able to recognize everyone’s style, their preferred openings, their preferred stroke sequences, etc.
Also see video:
The Canadian researchers therefore collected data from hundreds of thousands of games, cataloged by player. More than 1,000 parts for each of them, then broken down into sequences of 32 movements. Each move was then coded and fed to a neural network, such that each game represented a point in multi-dimensional space.
Thus, the moves of each player formed a set of points distant from the other sets. The system must then learn to recognize the differences between the moves and their sequences to distinguish styles. In any case, this is what the authors of the study explained during a presentation during the NeurIPS, a conference dedicated to systems that manage the processing of information in neural networks.
Once their algorithms had been trained, the researchers at the University of Toronto submitted 100 games conducted by 3,000 known players to the system, and then 100 games conducted by a mystery player. In order to make the task more complex, the first fifteen moves of the games were hidden. Thus, the artificial intelligence could not recognize a player by his openings.
The AI then went in search of the best game, and succeeded in identifying the mystery player in 86% of cases. A stunning result and far superior to an approach that did not rely on machine learning algorithms and was only 28% accurate.
Some researchers, notably Meta’s Noam Brown, who has developed a superhuman poker player bot, are delighted with these results, and are impatient to see the arrival of virtual chess players capable of playing in the style of grand masters, for example. They are even already considering other possibilities, such as seeing a bot learn to speak like a famous person.
However, a darker question arises. Because these algorithms, given the right data, would also be able to identify, with the same or similar degree of certainty, how a person surfs the Web. Then that would be the end of online anonymity, and it’s not hard to imagine all the abuse that could follow.
Moreover, the organizers of the NeurIPS conference found the study technically very impressive, but ethically dangerous. They agreed to this work being presented on the sole condition that the privacy risks were clearly discussed during the presentation. One of the peers who referred to this study also indicated that such a tool could be of interest “advertisers and sellers, or even the police force”.
One of the authors of the study indicated that he and his colleagues decided not to publish the code of their artificial intelligence. And too bad for the bot that could teach you how to play in the manner of the brilliant Magnus Carlsen.
Source : Science
The post This AI Is Sure to Identify Anonymous Chess Players…and Could One Day Track You Online appeared first on Archyworldys.