An AI can beat human players at chess, Go, poker, and other games that require multiple strategies to win. The artificial intelligence, called Student of Games (SoG), was created by Google DeepMind. The company says this is a step toward general artificial intelligence capable of performing any task with superhuman performance. A related paper was recently published in Science Advances.
Martin Schmid, who once worked on artificial intelligence research at DeepMind, now works at a start-up company called Equilibrium Technology. He said the SoG model can be traced back to two projects. One of them is DeepStack, an artificial intelligence developed by Schmid and other teams at the University of Alberta in Canada. It is the first artificial intelligence to defeat human professional players in a poker game. Another is DeepMind’s AlphaZero, which beats the best human players in games like chess and Go.
The difference between these two models is that one focuses on imperfect knowledge games - players do not know the status of other players, such as the hands in a poker game; the other focuses on perfect knowledge games, such as chess, where both players can see the position of all pieces at any time. The two require fundamentally different approaches. DeepMind hired the entire DeepStack team with the goal of building a model that could promote both types of games, and SoG was born.
Schmid said the SoG started as a "blueprint" for how to learn the game and then improve it through practice. This beginner model can then play freely in different games and teach itself how to play against another version of itself, learning new strategies and gradually becoming more capable. While DeepMind's previous AlphaZero could adapt to perfect knowledge games, SoG adapts to both perfect and imperfect knowledge games, making it more general.
The researchers tested SoG on chess, Go, poker, and a board game called Scotland Yard. They also tested SoG on Leduc poker and a customized version of Scotland Yard, and found that it could beat several existing AI models and human players. It should also be able to learn to play other games, Schmid said. "There's a lot of games you can just throw at it and it'll be really, really good at it."
This broad range of capabilities comes with a slight drop in performance compared to DeepMind's more specialized algorithms, but SoG easily beats the best human players in most games it learned. Schmid said SoG learned to play against itself in order to improve in the game, but also to explore what might be possible from the current state of the game, even if it was playing an imperfect knowledge game.
"When you're playing a game like poker, it's hard to figure out how to find the best next move if you don't know what cards your opponent holds," Schmid said. "So there are some ideas from AlphaZero, and some ideas from DeepStack, forming this huge combination of ideas, which is Student of the Game."
Michael Rovatsos of the University of Edinburgh in the United Kingdom, who was not involved in the study, said that although the research results are impressive, there is still a long way to go before artificial intelligence can be regarded as general intelligence, because games are an environment where all rules and behaviors are clearly defined, rather than the real world.
"The important thing to emphasize here is that this is a controlled, self-contained artificial environment where the meaning of everything and the consequences of every action are very clear," Rovatsos said. "This problem is a toy problem because although it may be very complex, it is not real."
Related paper information: https://doi.org/10.1126/sciadv.adg3256