Teaching AI to Play Cards and Games Is Not Just About Defeating Humans
Yesterday, DeepMind announced that its AI, AlphaStar, will log in to the European server and anonymously compete with human players in the StarCraft 2 ladder. Today, the news that the AI gambling god Pluribus, developed by Facebook and CMU, defeated the top human players in a six-player Texas Hold'em game has swept the screen again. AI is performing more and more prominently in game competitions, but what is the ultimate purpose and significance of our continuous training of AI to defeat humans in games?
Just yesterday, DeepMind announced that its AI AlphaStar will soon be available on the European server and will anonymously compete against human players in the ladder of StarCraft 2.Today, Facebook and CMU jointly developed AI Gambling God Pluribus , defeating top human players in a six-player Texas Hold'em tournament.
Why is AI always so interested in board games, and why is the team behind it working so hard to win games and board competitions?
Let’s start with the most complex strategy game “StarCraft 2”
StarCraft was launched by Blizzard Entertainment in 1998, and its sequel StarCraft 2 was released in 2010. It is considered to be the most popular game in recent years.The most difficult and hardcore real-time strategy game, while various games have been conquered by AI, it is a type of game that is relatively difficult for AI to master.
In order to win, players must carefully balance and handle multiple factors, and make timely plans and response strategies. Unlike chess games that rely solely on strategy, in this game, AI needs to face multiple challenges to play well, including dealing with imperfect information, making long-term plans, and learning timely strategies.
But last December, this situation changed completely. In 11 StarCraft 2 games between humans and AlphaStar, the AI won a crushing victory with a score of 10:1.At this point, AI has left a strong mark on the game of StarCraft.
AlphaStar's behavior is determined byDeep Neural NetworksThe neural network takes in the input data of the game interface (a list of units and their attributes) and outputs a series of instructions that constitute the in-game actions.
Based on imperfect information, games often last up to an hour and require tens of thousands of moves. Each frame of StarCraft is used as one step of input.The neural network predicts the expected sequence of actions for the rest of the frame after each frame and then takes the best action.
DeepMind explained that AlphaStar's success in the game was actually due to excellent macro and micro strategic decisions, rather than superior click-through rate or faster reaction time.
This technology can also help many other challenges in machine learning research, including modeling long-term sequences and large output spaces, such as translation, language modeling, and visual representation.
AI has taken over board games
- In 1997, the computer program "Deep Blue" defeated the world's top chess player at the time, marking the first time that AI has defeated human gamers.
- In May 2017, the growing AlphaGo defeated Ke Jie, the world's top Go player at the time, 3:0. Then, just five months later, DeepMind announced a new algorithm variant, namely AlphaGo Zero, was able to beat AlphaGo with a score of 100 to 0;
- At the end of 2018, Uber AI Research Institute used reinforcement learning algorithms Go-Explore,exist The Revenge of MontezumaThe score exceeded 2 million, with an average score of over 400,000, and is known as the strongest clearance algorithm in the history of Atari games;
- OpenAI Five First, we rely on 5 neural networks. OpenAI Five Defeat the amateur team of Dota 2 players. In April 2019, they defeated the world champion OG team 2:0 in the Dota2 International Invitational.
- Then, recently, the Texas Hold'em Poker God AI, jointly created by Facebook and CMU, Pluribus, in a six-player game, defeating the top human Texas Hold'em players can make an average profit of nearlyOne thousand dollars.
It took the team less than a week to train this Texas Hold'em gambling AI.
Humans have developed so many AIs that have defeated top human players in multiple complex e-sports games, making humans tremble in fear.
But does AI play games just for fun?First defeat humanity, then serve humanity
In order to teach algorithms to play Go, e-sports, and poker, these AI companies have spent a lot of effort and their attitude can be said to be very serious.
OpenAI even developed a Gym and Universe An open source platform that allows everyone to use this platform to teach computers to play games. Gym is used to play small games such as Atari, Flappy Bird, and Snake, while Universe is used to play large 3D games such as GTA5 and racing.
Do they spend so much energy and money just to play games for fun? Or do they feel a great sense of accomplishment when AI beats humans? No, for AI researchers, games are just a method, not a goal.
Game Environment:It is an accelerator for the overall development of AI
Games are the perfect testing platform for AI.Games have easy-to-process data, fixed rules, and a variety of hypothetical strategies. Games are simulated scenarios, which makes them an ideal place for the research and development of artificial intelligence.
“We haven’t seen a lot of things being trained in games that transfer to the real world,” said Julian Togelius, an associate professor at New York University’s Center for Game Studies. “But we have seen methods invented for playing games transfer to the real world.”
Game AI: The best teacher and opponent for human players
On the one hand, AI can help discover more perfect strategies and improve the competitive skills of human players. Ke Jie believes that playing against AlphaGo has opened his mind and improved his Go skills.
On the other hand, the intervention of AI will create a smarter opponent for many games. The adjusted AI can not only become a coach for human players, but also adapt to the levels of different players to play against them.
It can also prevent human players in today's games from getting angry at each other. If there is a Buddhist AI opposite you, it will definitely maintain a more civilized gaming environment. The game is just the beginning: there is a lot to do
DeepMind CEO Demis Hassabis said: "DeepMind's goal is not just to win the game, but also to have fun and be inspired by it."
“But personally, I like playing games, and I’ve developed computer games. But they’re also test beds in a way, where you try to write algorithms and test them.Ultimately, we hope to apply our technology to solve real-world problems."
In the future, AlphaGo and AlphaStar will not only be the names of AI players that control the heroes in the game, and DeepMind will not be limited to solving game problems. They will become AI heroes of human society.
-- over--