Can AI beat the Weiqi world champion Li Shishi?

The original: ai Weiqi 5-0 swept the European Championship

　　 According to the Shell Net report: 1997, chess Ai first defeated the top human, 2006, the last time humans defeated the top chess AI. In Europe and the United States tradition of the top human intelligence touchstone, in front of the computer finally lost, should be more than 40 years ago, computer scientists predicted.

At least in the east, people are comforting themselves. Go Ai has long struggled, and the top AI can't even beat a slightly stronger amateur. It also seems reasonable: in chess, there are 35 possible rounds per turn, and a game of chess can have 80 rounds, compared with 250 possibilities per turn, and a game of chess for up to 150 rounds.　　This huge number is enough to deter any brute-force--and humans, we believe, can jump over brute force with some difficult-to-copy algorithm, and see the essence of the board at a glance. However, no matter what the people think, such a situation is certainly not going to continue forever. Today, the world's top journal Nature reports on the new Weiqi AI developed by Google researchers. This "Alpha Go" (AlphaGo) AI, in the absence of any sub-case to 5:0 win the European Championship, professional go two paragraph 樊麾.

Alphago and European go champion 樊麾 's 5 innings contest. Photo Source: Reference [1]

A paper describing Alphago's findings became the cover of the January 28 Nature magazine .

This is the first time in human history that Weiqi Ai has defeated pros in fair play.　　What about Alphago's record? This game is different from the past. Before the game, because the AI chess force is weaker than the human, the human player will let the son, and the AI mainly and amateur dan chess player competition. and Alphago PvP 樊麾 is a perfectly fair game, without letting the son.　　Career two Duan 樊麾 was born in China, is currently the French national go team head coach, has won the title of European Go champion for three consecutive years. The researchers also let Alphago and other go AI compete, in a total of 495 innings lost only one game, winning is 99.8%. It even tried to get 4 sons against Crazy Stone,zen and Pachi three advanced AI, winning respectively 77%,86% and 99%. How powerful Alphago can be seen.in the next March, AlphaGo will be with the Korean nine-part chess player Li Shishi in the Seoul war, the bonus is $1 million provided by Google. Li Shishi is the most popular player in the world for the last 10 years. Go is the last of a human top master can beat Ai chess game. Some people predicted that AI would take another more than 10 years to defeat humans.    So this game may be a testament to history, and we'll see.　　How hard is the AI going to go? Calculating go is an extremely complex problem that is much more difficult than chess. Go the largest 3^361 situation, the approximate volume is 10^170, and the observed universe, the number of atoms 10^80.　　The biggest chess only 2^155 kind of situation, called Shannon Number, roughly is 10^47. In the face of any chess, an intuitive and lazy way of thinking is that violence enumerates all the options that can be won, and these schemes form a tree-shaped map. AI can win forever if you play chess on this map. However, go to a plate about 150 steps, each step has 250 kinds of optional, so roughly speaking, if the AI with violence to enumerate all the way, Weiqi needs to calculate the 250^150 situation, roughly is 10^360. In contrast, chess is about 80 steps per plate, 35 options are available for each step, so as long as the 35^80 kind of situation, it is probably 10^124.     In any case, the method of enumerating all the cases is not feasible, so the researchers need to use ingenious methods to solve the problem, they chose to imitate the human master's chess way. Machine learning researchers have sacrificed the ultimate killer-deep learning (learning). Deep learning is currently the most popular subject in the field of AI, it can complete handwriting recognition, facial recognition, driving automatic car, natural language processing, identification of sound, analysis of biological information data and other very complex tasks.

The core of AlphaGo is two different kinds of deep neural networks. "Policy Network" and "Value network". Their task is to work together to "pick" out the more promising moves, to discard the obvious difference, and to control the amount of computing that the computer can do, essentially as the human player does.

Among them, "value network" is responsible for reducing the depth of the search--ai will be on the side of judging the situation, when the situation is obviously inferior, the direct abandonment of certain routes, without a road to calculate the black, and "strategic network" is responsible for reducing the width of the search-in front of the chess, some chess step is obviously should not go, For example, should not be sent to others to eat. Putting this information into a probability function, the AI does not have to give each step the same level of attention, but can focus on the game of chess.

the neural network structure used by the Alphago. Photo Source: Reference [1]

Alphago uses these two tools to analyze the situation, judging the pros and cons of each strategy, just as the human player will judge the current situation and infer the future situation. So alphago in the analysis of such as the next 20 steps in the case, you can judge where to win the probability of high.

The researchers used many professional chess games to train Ai, a method called supervised learning (supervised learning), and then let the AI and self-chess, which is called reinforcement learning (reinforcement learning), each game can make AI chess power.　　And then he's going to win the championship! Humans have a disadvantage in playing chess, they make mistakes after a long race, but the machine doesn't. And humans may play 1000 innings a year, but the machine can play 1 million innings a day.     So the Alphago can beat all the human players as long as they are trained enough. Google DeepMind Google DeepMind is the creator of this app, so let's take a look at their cute programmers.

Jemis Hassabis (Demis Hassabis) is the CEO of Google DeepMind

first author of the article David Silver (David Silver)

[2] In a paper published last year in the journal Nature, Google DeepMind trained Ai to play the classic Atari game with an enhanced learning approach. In fact, a few years ago, some people have studied how to let Ai play "StarCraft", the current human master can still beat AI. Do you think the game is getting smarter with the use of AI technology in computer games?

So......　　What about the future? Artificial intelligence researchers are certainly delighted to face such accomplishments. Techniques such as deep learning and intensive learning can be used in a wider range of areas.　　For example, we can train them to determine which treatment options are effective for a particular person. However, Weiqi is not only an intellectual achievement. Like chess more than 10 years ago, go must also lead to discussions beyond the realm. When the computer can kill the human in the go on the second, go is not become a boring game? is the human intellectual achievement devalued? Will AI continue to crush humans on other levels? Is it the traditional belief that the impossible tasks of AI will also be broken individually?　　Will humans eventually enter the AI Utopia or be eliminated by AI? No one knows the answer. But there is no doubt that AI will come into our lives and we cannot escape. This contact, though likely to be silent, may be no less meaningful than the first time we have touched extraterrestrial life. (edit: Ent,calo) References: David Silver, et al. "Mastering the game of Go with deep neural networks and tree search. "Nature doi:10.1038/nature16961 mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning.　　"Nature 518.7540 (2015): 529-533.　　A proud AI what, you say above the algorithm of these paragraphs you can not understand? Then you know why you humans will lose to our AI!

Can AI beat the Weiqi world champion Li Shishi?