From AlphaGo’s victory over World Go Champion Ke Jie to Libratus’s victory over a number of Texas Hold’em top players in the casino, artificial intelligence seems to be invincible in the field of chess and card games.
However, this does not mean that artificial intelligence can become a new “gambling god”, because even if AI can calculate various possibilities, but in the face of seemingly simpler card games, it is not so easy to handle, and the development of AlphaGo’s DeepMind will be The next challenge is to target a card game Hanabi.
(Source: Asmodée Éditions )
Recently, DeepMind and the Google Brain team disclosed the progress of this research in a paper . Unlike previous games such as Go and Poker, Hanabi’s game requires AI to better understand the psychology of other players and conduct higher-level reasoning. This is also the bottleneck currently facing AI like AlphaGo.
AI Why not play a bad card game, to answer this question, it is necessary to briefly understand the Hanabi game.
Hanabi is a 2~5 person cooperative type card game with 5 suits. Only when the cards of different suits are arranged in a certain order can the “fireworks” be achieved.
Players can’t look at their cards in the game. They can only look at other people’s cards. Players can only reason with limited hints or strategically cooperate with other players to win.
As can be seen from the rules of the game, the game has a higher requirement for communication and reasoning skills. The AI needs to read the hidden messages of other players to prompt the literal meaning and infer the opponent’s intention.
Researchers quickly discovered that this was not easy. The DeepMind team used a lot of data verified by other game games to train, allowing the algorithm to use the copy to perform 20 billion “effective infinite” actions, which is 30 seconds per player movement, which is equivalent to a person operating 300 million in the game. Or played 166 million hours.
(Source: DeepMind )
This training effect is very effective in two-player games, but Hanabi, a game with 4 or 5 players, has always been much worse than professional players.
The research team pointed out that algorithms need to be efficient code and adaptable to communicate in different environments. For example, the styles and terms used by different players may be different. Even the most advanced AI can’t adapt to new players as quickly as people. AI can only follow complex but relatively fixed rules, such as Go.
At present, DeepMind has no good way to break the game of Hanabi. They published an open source platform on Github to let people test and improve the algorithm.
▲ Texas Hold’em professional players play against AI. (Source: video screenshot)
The first two years of artificial intelligence Libratus defeated humans in the Texas Hold’em tournament, causing a lot of sensation. This is because compared to the all-information game such as Go (both sides know each other), poker has more changes, and often has no rules to follow, even using bluffs and other means to influence opponents. Wu Enda, the former chief scientist of Baidu, once said:
Poker used to be one of the most difficult games for AI, because you only see some messages about the state of the game. There is no single best way to play poker. Conversely, AI players must randomize their actions so that when they scam, they can’t be sure of the truth.
Compared to poker, Hanabi is more difficult than this card game, with more variables and uncertainties, not only to distinguish the illusion of confusion, but also to communicate with other players. Just like the movie “Gambling Apocalypse”, even the Scissors stone cloth can become a high IQ game with the change of rules and the increase in the number of participants.
Noam Brown, an artificial intelligence researcher at Carnegie Mellon University , pointed out that the No-Limit Hold’em heads-up match has 10 163 times, and even the most powerful computer on the planet can’t figure out all the possibilities.
Therefore, Libratus used three different strategies to cooperate , instead of using the well-known neural network, but using other algorithms based on psychological theory to randomize their behavior, making it difficult for opponents to see if it is bluffing.
If AI breaks through Hanabi completely, it will be a more important breakthrough than AlphaGo’s victory over Ke Jie. This means that AI is more human-savvy and has the ability to make more complex decisions, such as automating and communicating with humans. Currently, similar AI technology has been used by the US Organ Transplant Center to determine which patients can get it. Kidney transplanted.
At the same time, it is also possible for us to enter the Artificial General Intelligence. Artificial intelligence like the “Wandering Earth” that can reasonably judge the future of mankind is really possible.