DeepMind introduced the Player of Games artificial intelligence system that can play poker, chess, Go and other games. The company has been investing in artificial intelligence gaming systems for a long time. DeepMind notes that games, although they have no obvious commercial use, demonstrate the cognitive abilities of artificial intelligence.
Player of Games, unlike other DeepMind gaming systems developed earlier (AlphaZero, AlphaStar), can work well both in games where strategy (chess) works, and in games where other factors are important ( poker).
In such games, the Player of Games learns to reason about the goals and motives of other participants, which paves the way for an AI that can successfully work with others, including solving issues requiring negotiation and achievement compromise.
“Player of Games learns to play games from scratch by simply playing on their own all the time,” said DeepMind Senior Research Fellow Martin Schmid, one of the co-creators of Player of Games. - “This is a step towards generality - Player of Games can play games with both perfect and imperfect information, while sacrificing some performance.”
While the Player of Games is extremely versatile, it cannot play any game. Schmid says that the system must take into account all possible perspectives of each player in a given game situation. Although in games with perfect information there is only one perspective, in games with imperfect information there can be many such positions - for example, about 2000 in poker. What's more, unlike MuZero, the successor to DeepMind AlphaZero, which picks rules for every game, Player of Games needs to be familiar with them.
In its research, DeepMind evaluated Player of Games trained using Google accelerator chipsets TPUv4, Chess, Go, Texas Hold'em and Scotland Yard strategy board game. For Go, a 200-game tournament was organized between AlphaZero and Player of Games, and for chess, DeepMind compared Player of Games to the best systems, including GnuGo, Pachi and Stockfish, as well as AlphaZero. The Player of Games Texas Hold'em match was played using the public Slumbot. In Scotland Yard, the algorithm played against a bot developed by Joseph Antonius Maria Neissen, nicknamed "PimBot" by the DeepMind co-authors. In chess and Go, the Player of Games proved to be superior to Stockfish and Pachi in certain, but not all configurations, and won 0.5% of games over AlphaZero. Despite these losses, DeepMind believes that Player of Games performed at the "best amateur" level and perhaps even at the professional level.
Player of Games was the best in poker and Scotland Yard.
Schmid believes the Player of Games is a big step towards truly general gaming systems, but far from the last. The general trend in the experiments was that the algorithm performed better with more computational resources (the player was trained on a dataset of 17 million "steps" or actions just to play Scotland Yard), and Schmid expects this approach to scale to foreseeable future.
AI experts estimate that AlphaZero costs tens of millions of dollars to train. DeepMind did not disclose the research budget for Player of Games, but it is unlikely to be lower.
The name of the Player of Games algorithm refers to the sci-fi novel by Scottish writer Ian M. Banks, published in 1988. It tells the story of Jernau Gurgeh, famous for his board game abilities.
Commentators point out that it would be interesting to test the algorithm in other card games.
Last month, DeepMind showed how it works. an artificial intelligence system helps mathematicians in finding information to develop theorems. The collaborative work of researchers and AI has already led to a breakthrough in hypothesis in the field of topology and representation theory, as well as a proven theorem on the structure of nodes.
In October, DeepMind reported for the first time about profitability. The company ended 2020 with £ 43.8 million ($ 59.6 million) in profit.
bbabo.Net