Original address: https://bunshun.jp/articles/-/43711
Original title: いま囲界できている The relationship between the human world and AI - "The era of China's top two enterprises" The Strongest へと Growth" Shogi World とは一なるAI との Xiang き合いかた
Excerpt from: Wenchun Online
Pictures: ITMedia, h-eba
Reporter: Shiro Shiro
Column strives to end this month
Shirou Shirou: So, does Go AI use percentage to represent the situation? Shogi AI uses points.
Bridge extension: In Go, the situation is expressed in a way such as "black winning 55%". But recently I feel this number is also very interesting. It is the star array that displays the mesh number below the winning rate. I feel that this combination makes the star array stronger and stronger.
Shirou Shirou: Are the two numbers together?
Bridge extension: Yes, because Go is a game that sees who has more territory, so even if the previous AI said that the black chess winning rate was 90%, it still did not know whether it won 50 games or 1 game.
If the winning rate is 90%, it is only 1 point ahead, then a small mistake can be reversed. But AI doesn't make many mistakes, so there is a 90% win rate, but humans often make mistakes, so at this time, it's not a 90% win rate at all. Then the star array will tell us that 90% of the blacks here are black leading by 1 eye, telling us that the winning rate is closer to human beings.
And recently, KataGo's Go AI has appeared. This AI is also a free and open source software , and many developers have transformed it. Like the software "Yaneura King" in the shogi world.
If we only talk about open source software, Leela Zero was the most popular until last year, but now I feel that KataGo is even more popular because KataGo can display mesh numbers.
Shiratori Shirou: Will it help you to show how much data you are ahead of?
Bridge extension: yes. Star Array and KataGo are popular recently, but for open source software, KataGo is better.
Shiro Shirou: If the mesh number can be calculated, that means the weakest end-game capability in deep learning has been strengthened?
Bridge extension: Yes, deep learning itself is best at recognizing images. Go is like stippling, deep learning can identify whether to win or not from the position. But it needs to calculate the number of meshes, so you need to search for more things. Of course, it is really difficult to let the AI learn this technology at the beginning, but once you get on the right track, you gradually feel that this will make the AI stronger.
However, it cannot be said that it completely depicts the thinking of human beings. Now our goal is to make AI more human-like.
Shiro Shirou: So, after AlphaGo came out, young players looked at the winning percentage and thought, now they think according to the number of goals?
Bridge extension: This is a very complicated place. Like KataGo and Star Array, the winning percentage and the number of meshes are displayed on the chessboard to judge the situation.
However, the number of meshes is very subtle, and the difference between the first selection point and the second selection point may be only 0.1 mesh.
Shirou Shirou: It turns out that it is not a difference of 1 order, but a difference below the decimal point.
Bridge extension: Sometimes the gap between the first selection point and the tenth selection point is one point away. This is also common.
Shiro Shiro: Wow, I really don't know where to go.
Bridge extension: Specifically, I look at the winning percentage in the layout stage, and then look at the number of meshes in the final game. In the layout stage of
, sometimes the first selection point and the seventh selection point are only one game away. It is really difficult for us to understand these changes (laughs).
However, if the winning rate is 7%, for example, 49% and 56%, is it a lot worse?
Shiro Shirou: The computer tells us the winning percentage and the number of meshes. Has the learning efficiency of human beings improved?
Bridge extension: After the number of meshes is still available, the learning efficiency of chess players has been greatly improved. Sometimes AI shows that the winning rate is 70%, but winning 10 goals is 70%, and the difference is only 70% with 1 goal. We have a lot of things that we don't understand in this regard.
But sometimes "how come it's 70% with only 1 game here?", or "I've played so many chess, how can the winning rate be 70%" and so on.
Shirou Shirou: In this case, how about the 70% win rate?
Bridge extension: It depends on the situation. Sometimes even if you can take a lot of chess, then you can win all the chess if you are right and capture your opponent, but sometimes even if you make a wrong move, the winning rate will drop significantly.
Shiro Shirou: So, even if it is only one eye away, sometimes you don't have to go to great lengths and you can win by running it safely. This is very similar to shogi.
Bridge extension: This is the "dystopia of winning percentage" (laughs). Even shogi software can become stronger by combining numbers and percentages.
If the number of Go games is compared with the shogi idea, for example, an AI can be created that uses the shortest order to close.
Shiro Shirou: That's right.
Bridge extension: Also, the deep learning software of the Zero series requires a lot of money. So I hope everyone who develops shogi AI can find a more efficient method, and then it would be best if it can be used in the field of Go (laughs).
Shiratori Shirou: Funding is, how much did you invest in development?
Bridge extension: The DeepMind team that developed AlphaGo borrowed the server from Google for development, and then spent 35 million US dollars. I was shocked to see this news.
Shirou Shirou: Huh? One US dollar is converted into about 110 yen, which is about 4 billion yen? ah? Did you spend so much money to develop deep learning?
Bridge extension: In short, huge resources are needed. Taking AlphaGo as an example, the TPU used 2000 bases, and Facebook also used 2000 bases of GPU when developing ElF Open Go. When we developed GLOBIS-AQZ, we also used 1000 bases.
Shiro Shiro: 1000! Do you need so much?
Bridge extension: 1000 bases, which is basically impossible in and in Japan.
Shiro Shirou: So do you want to use an overseas server for development?
Bridge extension: No, we use the large-scale AI cloud computing program "ABCI" of the Industrial Technology Research Institute.
Shiro Shiro: ABCI of the Institute of Industrial Technology, got it. In 2018, it is still a large-scale cloud computing system with the world's top five performance. There is a 4352-based high-performance GPU. It turns out that such a system is also available in Japan.
Bridge extension: There was the best GPU over 4000 bases at that time - V100, we used it for about a year. We used 1000 bases the most and really appreciate them.
But the development of GLOBIS-AQZ has now also stopped. Large-scale development is no longer carried out by enterprises as a unit, and it is like chess that pin its hopes on individual developers.
later asked several developers, we all had a good time chatting about technology, but in the end we always asked a question: "Is it okay for Japan to go on like this" (wry smile).
Shiro Shirou: But the unique art developed just like this is also very strong. How many resources did they invest? It can reflect that China's IT industry has been far ahead a lot.
Bridge extension: Go is roughly divided into Japanese rules and Chinese rules. GLOBIS-AQZ is developed according to the black sticker 6 and a half.
But in , Europe, America, and countries, like Douyin, Chinese software is banned from entering the country, so even if they can use their Go AI abroad, maybe one day they will not be able to use it.
Shiro Shiro: In order to avoid this situation, you need to develop in your own country, but this requires a lot of resources, which is indeed a very complicated problem.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-