The release of AlphaGo is a great milestone, which makes me excited to talk about it again.

Let me start with my assertion that 1: AlphaGo will completely defeat Li Shishi in two months.

Leave a mark. At the end of this article, I will talk about assertion 2.

Since middle school, I have been fascinated with solving game problems with algorithms, and I have creatively completed some problems with search methods. After the emergence of deep learning in recent years, I feel that I have a chance to break through Go. After several discussions with the joint laboratory in Tsinghua, I think this direction is feasible. Unfortunately, due to the lack of gas punishment ability, it is impossible to organize investment in this area.

This time, Google's DeepMind team is the top team in the field of deep learning. There are no problems in resources, abilities and gas fields. The breakthrough technology is based on deep learning of valuation and chess.

After watching a lot of discussions on Zhihu, we can judge AlphaGo's chess style from previous chess games and infer that this algorithm is powerful, which is a bit like carving a boat for a sword. Our core is to return to this in-depth understanding of the technology used by AlphaGo. In order to facilitate discussion, we compare three differences between Deep Blue Chess with search pruning as the core and AlphaGo with search pruning+deep learning as the core:

1. The biggest difference between Go and chess is that the evaluation function of chess is extremely difficult to define. Chess can find various "characteristics" to score points, such as how many points a horse loses, and how many points a soldier bows forward and approaches the baseline, but Go can't. Black and white players are so dense that they are related before, so it is difficult to sum up the rules. This is also one of the weakest problems of traditional algorithms. Just as we do face recognition, we can know at a glance that it is Zhang Lisan, and the machine algorithm is difficult to start. This problem is precisely the biggest breakthrough in deep learning in recent years. Deep learning does not require people to design algorithms to "find features". Through a large number of original data and tag pairs, the machine can automatically find features, and it is no worse than people. A few years ago, many people thought that machines were struggling in image processing. How to define and abstract the nose? Ears? Eyes? But in the past two years, deep learning has advanced by leaps and bounds, surpassing human beings in one fell swoop. 20 15 in face recognition, the recognition ability of machines has surpassed that of people, which is one of the core abilities that people have evolved for tens of millions of years. A game of Go can be understood as a picture of 19* 19, and the rules of other games are very simple (easily translated into computer rules), which just falls in the place where deep learning is good at. Search+deep learning, this algorithm can completely cover the rules of Go, and the thinking process and mode of people playing chess is only a subset of AlphaGo. This determines that this algorithm has no ceiling and has the opportunity to "clear customs" in the field of Go.

2. Compared with AlphaGo, the biggest advantage of deep blue is "learning ability". The opening of Deep Blue depends more on the establishment of database chess, but it has no generalization ability (it doesn't know how to draw inferences from others), so it may be silly to play chess in an unprecedented way. After that, the core ability is computing ability. Through violent search (of course, there are the best pruning, but it is still violent), I try to go out of the 10-20 step to choose the best path. This kind of complexity is exponential and becomes NP problem, which is limited by computing power. The algorithm of this system is written to death, and it will have fixed performance under fixed parameters. Adjusting parameters and changing algorithms are all things for engineers. The ceiling of this system is how powerful the computer is and how smart the engineer is. AlphaGo is more data-driven. By feeding him more chess data, he can optimize the "neural network", become smarter with the same computing resources, have the ability to draw inferences from others, and be very close to people (or designed by simulated people). Moreover, we know that the machine's ability to process data is fast enough and there is no mood to make mistakes, which determines that if this system learns all the chess games that can be collected on the Internet today, it will become a top chess player.

The most terrible thing is not just the above two points. For chess, AlphaGo not only collects data from the Internet for learning, but also can learn by itself and do it by itself. Have you seen the movie Transcendental Hacker? Over time, artificial intelligence can become smarter. In Jin Yong's novels, the old urchin made his left hand and right hand "hit left and right" and became an invincible martial art in the world. That's just a story, in the field of chess, and AlphaGo has such a design to make this martial arts come true! There are still two months left. AlphaGo, a machine with no ceiling, has a good chance to reach its peak and become an invincible Go master.

AlphaGo's technical problems are over. How to see the complete action behind Google? Some people think that it is over-interpretation, and the actual system is quite rough-all choose "European champions"-indicating that the system is not good, which is a wrong understanding. The more likely reason is that Google and Facebook are competing to play Go, and Facebook employees are careless and leaked a lot of news in advance. As a result, Google quickly grabbed the article of Nature, and then sold a futures for two months to compete with human beings. At that time, the system was good enough, which was an appropriate practice in a competitive environment.

In fact, both Google and Facebook recognize the importance of artificial intelligence and will make great breakthroughs in recent years. Google bought DeepMind for $400 million. There were only 20 people at that time, but now there are more than 200 people, and they are crazy at all costs. Playing Go is only an excellent propaganda point and breakthrough point to reflect the progress of artificial intelligence. As can be seen from the public literature, DeepMind's research and development of Go is based on general technology and independent of the field. This technology can be used in other suitable fields in the future. The charm of deep learning is that as long as you can model in a field and have enough data, you can surpass and replace people in this field and get 99 points from 0 in a short time. If we are still conservative, we should understand machine intelligence step by step. For example, a big boss publicizes that his XX brain has achieved the intelligence of X years old, which is very misleading. We will also misjudge the ability of Go machine and evaluate it according to human understanding of 1D-9D. In short, don't use the method of evaluating people to evaluate the artificial intelligence ability of machines. This is a completely different model.

Lao Luo once commented on a sentence of artificial intelligence: "Artificial intelligence is like a train. As it approaches, you hear a rumbling sound, and you are constantly looking forward to its arrival. He finally came, flashed by and left you far behind. "

If we put a patch on this sentence and limit the application of artificial intelligence to a specific closed field, this is a very appropriate description. We should not be too conceited. For example, driven by our sense of self-superiority, we tend to say that animals are inferior to people. For example, people can walk upright, talk and use practical tools to distinguish themselves from other animals. Facts have proved that animals can do the same. The same is true of machines. Just a few months ago, someone clamored that the machine could not play Go for ten years. The reason is that people can understand it at a glance, while machines can only calculate. These self-assemblies make us misjudge. Don't feel too inferior. I feel that the whole intelligence of human beings has been crushed by the victory of Go. Today, there are still many fields where machines are completely incompetent, only in local areas.

Finally, let's talk about assertion 2: In addition to Go, artificial intelligence will sweep everything and completely defeat human beings in other closed games.

Although Wang Xiaochuan's statement is too arbitrary, the possibility of AlphaGo winning is not low, and its real level of Go may be far more than it is now. First of all, the match between AlphaGo and Fan Hui actually happened at 10 last year, and it was only recently exposed. We don't know how much AlphaGo has improved in the next few months. Secondly, after this incident was reported, a machine suspected of AlphaGo appeared on the Internet of Yicheng Chess Institute, where experts gathered. Judging from its number of battles and results, it has fought many battles with human beings. Even, the best score reached 9 days (segments). It can be seen that AlphaGo has been hiding its edge. As Wang Xiaochuan said, Google must be confident enough to let AlphaGo challenge Li Shishi.

Of course, it is still too early to say whether AlphaGo can win the game at the end of next month. However, this will definitely not be an unstoppable game, but the peak confrontation between the masters.