The new Alphago, which had a 60/0 record against professional go players in online fast chess, beat Ke Jie, the world’s first go player, in a slow chess match yesterday. How powerful is ARTIFICIAL Intelligence? Has Go been “conquered”? Is there hope for humanity?

Zheng Yu, a staff writer and expert on artificial intelligence and big data at thepaper.cn, explained the above questions on May 24:

Five myths: Humans aren’t that bad, alphago won’t lose on purpose

Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include information robot, editing robot, writing robot and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.

To this situation, Ke Jie continued to grab hair for 25 seconds, leading to the scene of a joke

On May 23, the version of AlphaGo2.0 defeated Ke Jie in the man-machine Go match by a narrow margin of one quarter of a son, which on the one hand continued to demonstrate the power of artificial intelligence, but also let people have a new understanding of AlphaGo and artificial intelligence.

Conclusion first, analysis second:

1. AlphaGo is currently ahead of humans in the game of Go, but it has not completely conquered the game. It just uses deep learning to find a better solution than humans know, but it’s not optimal. The optimal solution cannot be found, even with all the resources on earth. From a professional point of view, it is the method of using deep learning to approximate a value judgment function in reinforcement learning, and then combining it with monte Carlo search tree (see Zheng Yu: A Diagram of ALphaGo Principles and Weaknesses, not to repeat). Since neither AI nor humans can find an optimal solution, it is too early to say which side has completely and utterly failed.

2. Human beings are also making progress. We should not underestimate the rapid (small sample) learning ability of human beings, which AlphaGo cannot do based on the current learning method. In the short term the probability of winning is low, but in the long term (the next 5-10 years) there is a chance because people are also very good at learning and can learn quickly from a small number of games played against AlphaGo. Even if AlphaGo is given another 100 million games and another 10,000 Gpus, its progress will slow down if it is based on the existing learning system, because the new games and computing resources are just a drop in the ocean compared to the 2×10171 search space. We don’t know nearly as much about the human brain as we do about Go, and there are big unknowns.

Ke jie’s first two moves, “3 ·3”, have overturned at least two decades of weiqi theory (from weike social App)

But ke’s move is not a whim. He has used it many times against other professional masters in various professional matches. In the recent Professional competition of “biao Jia” in China, various “dog moves” (Alphago moves) have also emerged in an endless flow, and people are trying to use their own understanding of the moves created by artificial intelligence. And professional go players such as Ke jie’s “feeding and removing moves” are also an important factor in the rapid growth of China’s own Go AI “Ju Yi”.)

In fact, Ke has tried this trick several times against humans (from weike, a social go App)

Li Zhe, a “scholastic go player” now majoring in philosophy at Peking University, also tried a variety of new moves.

3. At present, the gap between human professional chess players and AlphaGo is only at the same level, not as big as everyone imagined. In fact, this gap (according to the Chinese standard of 7.5 mesh), in the eyes of professional chess players, is already a very large gap. A lot of professional ace, after entering official child stage discover oneself to still lag the other side 7-8 eye, can throw in the towel actively. Many games decided by counting pieces were won by one or two mesh lines (Ke jie lost to AlphaGo by half a mesh, for example). Otherwise, they will be laughed at by other professional chess players. They don’t know how much they are behind, and their ability to point air is too weak.

To be able to view this problem objectively and accurately, it is urgent to have strong professional knowledge of artificial intelligence, as well as certain go skills. The following is to correct some misunderstandings of online cognition:

Myth # 1: AlphaGo allows top human players to have 4 pieces. AlphaGo2.0 allows the previous version to have 4-5 pieces.

To dispel this misconception, it is important to popularize the knowledge of Go: there is a world of difference between giving your opponent two pieces and winning two pieces. This goes without saying to go players, but I realized today that many melon eaters have always thought this was the same thing. No wonder the above erroneous remarks spread online.

Let the other two pieces:

To make two pieces in Go is to ask one side to put two pieces on the board (the pieces can only be placed in star position) before the other side starts to move. The value of these two pieces is huge at the beginning of the game, and for professional players, each piece is worth at least 10 pieces (this is the most conservative estimate). Let two sons equivalent to at least first let out each other more than 20 purposes of turf. Since the board is limited, if you can’t win back the 20 or so pieces in the limited space later in the game, you have failed. Moreover, the more the number of suitors, the value obtained by the transferee will not only increase linearly, because there will be cooperation between the suitors to obtain greater benefits. Rangzi, for example, could be worth much more than 40 mesh.

Win 2 pieces: refers to the two sides played, the winner than the loser of the party more than 2 pieces. If according to the method of eating and housing each other, the two pieces are only equivalent to 4 eyes. AlphaGo won ke Jie by a quarter of a piece, which is equivalent to half a piece.

So “give the other party two pieces” and “win the other party two pieces” can not be the same language. If there is such a thing as a Go god (and since he is sure to find an optimal solution, we have no hope of beating him), the average top professional player thinks they are within two or three pieces of this god. Since AlphaGo can be shown not to be guaranteed to find the optimal solution, he is still some way from god. Therefore, it is impossible to say that AlphaGo can make the top human chess player four sizes.

Myth 2: AlphaGo also plays some obviously bad moves because it relaxes its own requirements in its own judgment of form.

AlphaGo’s search strategy, which prioritizes more deep searches on branches with higher probability of winning, will not change at any time and cannot be changed. He’s not gonna think he’s got an advantage and go easy. When AlphaGo plays poorly, it is because its value judgment is an approximation, and the search space is not exhausted, and the optimal solution is not available. Therefore, sometimes the estimated good moves are not necessarily the best moves, and it is normal for AlphaGo to have such an unstable situation. This is also the hope of human existence. Of course, human beings also have their own weaknesses, such as fatigue, mood swings and so on, people can also make mistakes in judgment. Moreover, the board is long, and some of the previously bad moves can become good ones through subsequent changes (including unexpected changes). So, not all mistakes will directly affect the outcome of the game. And now everyone seems to be a little afraid of AlphaGo, even if AlphaGo makes a bad move, everyone is more skeptical of their own level (is it that we did not understand ah?). “And chose to trust AlphaGo’s” foresight “.

Myth 3: AlphaGo can constantly learn from itself, gaining experience in new games and improving itself quickly.

Because AlphaGo’s system has so many parameters, it needs a lot of data to train, and the addition of a few game charts will do nothing to improve its performance. Moreover, when AlphaGo adjusts its parameters, it aims at the overall optimization of a large number of data and has to process many chess scores in batches. The training time is very long, so it is impossible for AlphaGo to significantly improve its own level in a very short time. Even if the same group of trained chess score, different parameter adjustment methods will train the chess ability level difference system. In fact, AlphaGo generates scores of games by playing against itself, and then trains a value network using the relationship between the (two consecutive) sides of the score and the final winners. It just borrows the framework of reinforcement learning to train the parameters of deep neural network, and its main contribution is the approximate ability of deep learning (it solves the difficult problem that traditional reinforcement learning cannot solve for complex environment and action state). As a result, AlphaGo does not have the ability to make progress on its own by playing against itself as many people think.

Myth 4: AlphaGo will lose a game on purpose.

That’s impossible. To lose without losing so ugly and obviously is a very difficult thing to do, perhaps even harder than winning chess. After the model is trained, the only thing AlphaGo can change temporarily is the amount of resources (or space) it spends on the search. It’s too small to make a big difference, but if it’s too small, it can make some very low moves. This is technically difficult to grasp.

Myth 5: The computing power of a computer must be stronger than a human, so don’t compare with AlphaGo. You should simplify the situation and avoid complex battles.

AlphaGo relies on a tree-based search algorithm, making it harder to judge the value of future wins and losses as the search space expands in complex situations. As a result, complex situations beyond human calculation are also difficult for AlphaGo. If the situation is too simple, machines can do a better job, and human players have even less hope. So by making things complicated, a human player can hope to win, even if it presents a greater challenge to humans.

Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.

conclusion

Based on the current knowledge and understanding of Go, humans will still lose to AI at this stage. I don’t think Ke jie has a chance to win the following matches, but human beings are also making progress. Through playing AlphaGo, human beings are also relearning the concept of Go. As long as the human mind and civilization continue to advance, it is possible to catch up with the current AlphaGo through constant learning in the next 5-10 years. AlphaGo will improve, of course, but it is not the God of Go, nor has it cracked the game. Without a comprehensive overhaul of existing learning methods, the rate of progress will slow down. With that in mind, there is still a chance. As people gain a better understanding of Go, better artificial intelligence algorithms will be devised. In fact, the two are not contradictory. They complement each other and reinforce each other. No matter who wins or loses, it is a reflection of the progress of human civilization. Human intelligence will always be ahead of machines and will not be replaced by machines.