The official partner of the National Go Team, Junior Zongheng, made a special report.
After Ke Jie burst into tears in Wuzhen in 2017, the confrontation between human Go and AI Go was completely over. Although AlphaGo has retired, according to their relevant papers, many Go AI such as Jueji, Tianran, StarArray, KataGo and so on have been derived, some of which have become tools for people to study Go everyday. In the eyes of experts who develop these AI, what is the question of "has AI conquered Go?"?
"Man Machine War" between Tian Ran and Park Tinghuan
Li Kaihua, the former head of Tianyan Weiqi, defeated Park Tinghuan in human-computer game in 2018.
Sina Chess: Do you think AI has conquered Go?
Li Kaihua: No. The number of changes in Go determines that its optimal solution set may be an inestimable size. AI certainly didn't conquer Go. What AI has achieved so far is that the number of Go changes it has explored is many orders of magnitude larger than the number of Go changes that humans have explored, about: all the training chess manuals of AI at present/(all the chess manuals that humans have played+the change map that humans have studied)=about 10000 to 100000 times.
There is still a big gap between this number and the total number of legal changes in Go, which is 10 to the power of 170.
Of course, for scientific research, AI has achieved the maximum exploration under the existing technical conditions. There are two reasons:
1。 Through mathematical deduction, it can be concluded that the chess power can be constantly strengthened through deep neural network and reinforcement learning. Therefore, as long as human beings have enough time and enough storage devices to generate and store 10 170 power disks, they can certainly find the optimal solution of Go. It is OK to follow this path in scientific research. After all, even if every atom can have a legal change, there are only 10 to the 70th power atoms in the universe, which is 10 to the 100th power times more than the total number of legal plates of Go.
2。 Go is a zero sum game with complete information. In contrast, the non zero sum game with incomplete information such as Delphi and Mahjong, as well as the non zero sum game with cooperation, has even jumped out of the game theory level. At the generation level, there are more complex situations that need to be studied. So in 2014, Ian Goodfellow developed a generative confrontation network, In 2017, Google Brain came up with Attention is All You Need, which is the basic Transformer model of ChatGPT. Then the forefront of AI research has always been in the Transformer model.
3。 Now, the conclusion is that objectively AI has not really conquered Go, but in principle it has found a way to conquer Go, so it can be considered scientifically that "AI has conquered Go".
Sina Chess: To what extent do you think the highest level of AI Go has reached?
Li Kaihua: 1. Because the legal number of Go is 10 to 170, humans have explored no more than 1 million changes, and AI has explored almost 10 billion changes. Even so, there are 10 to the 160th power changes, so if the Go God is 100, the AI has explored less than one trillion. Human is one ten thousandth of AI. It is far from the previous 7%.
2。 According to the latest KataGo reinforcement learning, the latest model of AI has reached about 13500 Elo points.
3。 Note that Elo score can only be used to compare the strength of each other under the same system, that is, the Elo score of AI and the grade score of people are two systems, and cannot directly compare values. Unless an anchor point is found, for example, Li Shishi's Elo score is used to estimate the chess power of the AlphaGo Lee version.
4。 However, the strength of Elo points can be effective only after two players have played enough games. For example, Gulee and Li Shishi played dozens of chess games. The difference of Elo points between them can indicate their strength, but if the error of Elo points is within 5 points, the two players need to play at least 89 games. If the error is within 3 points, two players need to play at least 200 games. Therefore, it is purely entertainment to use Li Shishi's Elo score to bring in the Elo score of AlphaGo and compare the chess power gap between AlphaGo and the chess players at that time. Too few games.
5。 Elo scores also expand, that is, if the models are all of the same style or system, and there is no interference from external models, if the models at one stage are over fitted, the scores will soon expand because of these models. It cannot be corrected subsequently.
6。 At present, if we roughly estimate the chess power, it is possible for the strongest KataGo to have 3-4 chess players with a good distributed server, such as 100 A800 graphics cards or the latest GB800 graphics card. But I'm afraid the electricity cost of that chess game will be 100000 yuan less.
Park Tinghuan's Ninth Duan in 2018
Sina Chess and Card: Where is the end point of the follow-up research? Is there any difficulty?
Li Kaihua: 1. The route from AlphaGo to Leela to KataGo is equivalent to letting a baby learn to play chess only by looking at the chess manual. The brain can only play chess. So a lot of chess manuals are needed to teach him. At the same time, because he is actually learning to play chess by looking at pictures, he has no logical reasoning ability, so if he wants to become stronger, he can only play chess by himself constantly and see more changes to learn to play chess. It looks heavy.
2。 In the future, we may hope that ChatGPT, a kid in his teens who has an IQ but knows everything, but has certain linguistic reasoning ability, can learn to play chess by studying chess scores instead of looking at pictures, and can use less chess scores to achieve the same intensity as AlphaGo. Even the stronger version of ChatGPT 5.0, if it can be equivalent to an adult's reasoning ability, should be more smooth.
Man machine game scene
Sina Chess: What is the development of AI Go sparring technology?
Li Kaihua: 1. If the language model can be used to play chess, it is easy for the language model to explain why it plays chess so well.
2。 At present, professional chess players mainly use backrests and guesses to understand chess under artificial intelligence.
3。 The level and difficulty of AI sparring for children can be grasped very well, because the improvement of children's chess ability is discontinuous, and it is often a case of phased leap. In addition, the computing power and game ability corresponding to different chess power are not improved linearly, which leads to the fact that all current tuning parameters are stupid and basically no parameters can be adjusted. This aspect may be a long-term problem. It's better to wait for 4 (a) to be solved and then hand it over to AI.
(Tour around)
Statement: Sina.com exclusive manuscript, unauthorized reproduction is prohibited!
