AI can beat various world champions, but becomes an "amateur" when playing cards!

In 1994, an AI called Chinook defeated humans in the World Checkers Championship.

In 1996, IBM's Deep Blue defeated world chess champion Garry Kasparov for the first time.

In 2007, researchers at the University of Alberta built a checkers-playing computer program that was unbeatable.

In 2011, an AI named Watson defeated its human opponents on a quiz show.

In 2013, DeepMind took on Atari and won.

In 2016, AlphaGo defeated Go grandmaster Lee Sedol.

The match between AlphaGo and Lee Sedol.

AI is getting more and more powerful. It seems that humans cannot even compete with AI in this board game. When we look back at the history of AI, we find that the inspiration for the breakthrough in artificial intelligence itself comes from the board game of checkers .

Machine learning is an algorithm in AI that allows computers to learn on their own, rather than relying on manually coded solutions by humans, to become "smarter". In the 1950s, Arthur Samuel explained machine learning in his research on the game of checkers. He mentioned that there are two ways to make computers better without human intervention: the first is rote learning, and the second is groundbreaking generalization learning. This method allows the program to know where to go and review previous game rounds.

Soon after, Arthur Samuel designed a checkers-playing program at IBM that could easily beat novice players, and then, also from IBM, designed an AI that could play chess.

Arthur Samuel's checkers program showing it. References [1]

Later, when we talked about AI, we could see that it could far surpass excellent human players in many games. However, in 2019, AI lost once, and the reason for the loss was really "stupid".

At that time, DeepMind tried to use AI to challenge a card game called Hanabi, which has 5 sets of cards. 2 to 5 people work together. During the game, players cannot look at their own cards, but can only look at other people's cards. Through reasoning and some cooperation, players need to arrange cards of different suits in a certain order.

It is this card game called Fireworks that stumped the AI. Image source: wiki

It seems that this is a very simple game. Players over 8 years old can play it, and the game time does not exceed 30 minutes. The key points of the game are communication, reasoning, and cooperation with others.

However, the simple game in the human world gave AI a little shock. Researchers at MIT Lincoln Laboratory used advanced AI models to play the game, and found that the AI ​​did not perform well.

This game requires cooperation, but when people and AI play together, people find that AI is so unreliable and unpredictable that people lose trust in it. Even if they cooperate well, players still feel negative.

AI has always performed worse than chess in various card games. In 2017, scientists made AI Libratus beat humans in Texas Hold'em, and in 2022, AI achieved a milestone: beating the bridge champion.

Some argue that there is one thing that makes poker unique: it is very human. Famous bridge player Warren Buffett once said, "Playing bridge is like running a business. It's a game of hunting, chasing, nuance, deception, reward, danger, and cooperation."

AI needs to be more human-like to play card games well. Image source: Pixabay

In 2022, AI was able to beat human bridge players because it made AI's learning method closer to that of humans. That is, AI learned to explain how a decision was made. Hanabi may seem simple, but it actually requires reasoning and explaining the intentions of other players.

References

[1]https://www.andreykurenkov.com/writing/ai/a-brief-history-of-game-ai/

[2]https://arxiv.org/abs/2107.07630

[3]https://www.wsj.com/articles/why-the-card-game-hanabi-is-the-next-big-hurdle-for-artificial-intelligence-11553875351

[4]https://www.businessinsider.com/ai-vs-humans-in-different-games-2016-11

[5]https://www.theguardian.com/technology/2022/mar/29/artificial-intelligence-beats-eight-world-champions-at-bridge

[6]https://www.technologyreview.com/2017/01/23/154433/why-poker-is-a-big-deal-for-artificial-intelligence/

[7]https://www.technologyreview.com/2017/01/11/154629/poker-is-the-latest-game-to-fold-against-artificial-intelligence/

Planning and production

Source: Bringing Science Home

Editor: Yinuo