It is currently Thu Mar 28, 2024 3:47 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 2 posts ] 
Author Message
Offline
 Post subject: Interview with Golaxy developer Dr. Jinxing
Post #1 Posted: Wed Aug 03, 2022 12:25 am 
Beginner

Posts: 11
Liked others: 0
Was liked: 2
Rank: Chinese amateur 5 Da
KGS: 1 Dan
Tygem: 3 Dan
IGS: 4 Dan
Wbaduk: 4 Dan
OGS: 7 Dan
Starting with Golaxy ——An interview with Dr. Jin Xing, the developer of Golaxy Interviewed and compiled by our reporter
Thematic Research

May 2019

Golaxy is a Go-playing program
launched in China in 2018, and its predecessor was the Go AI "Shensuanzi" (Abacus in English) developed by Tsinghua University, which was later taken over by the Beijing-based company Thinker Technology for development. Golaxy rose to fame after winning AI vs. human Go games against Ke Jie, one of the world's best Go players, in Fuzhou in April 2018. Golaxy has been very active since its debut and has maintained a unique stance among AIs with its "non-compromising Go". Golaxy has played more than 70 open games against professional players in 2018 with few losses and has won the 2018 CITIC Securities Cup World Go AI Competition and the World Go AI RYUSEI. On April 29, Golaxy defeated Korea's BaduGi 3-0 in the final of the 2019 World AI Go Game in Fuzhou.

Golaxy's developer, Dr. Jin Xing, is a technical expert in the field of AI and has unique insights into the development and innovation of Go AI. With the increasing interaction between AI and the Go industry, people's perception of Go AI is gradually transforming from a sense of unfamiliarity with an exotic object to an acceptance of something new. Of course, we are very interested in discussing various topics of Go AI with professionals, so we invited Dr. Jin to talk about his story and his program at his company located near Tsinghua University in Beijing.




Q: Congratulations to Golaxy for winning the championship as expected! The competition went quite well. The three games in the final were almost overwhelming. What do you think of Golaxy's performance in this competition?

Dr. Jin: Thank you! It did go really well. The data we saw in the background showed that Golaxy's win rate did not fluctuate too much, and the win rate in the final three games was steadily increasing. Once the AI's winning rate fluctuates greatly, it will show a cliff-like drop indicating miscalculation or even serious miscalculation, so sometimes a sudden increase in the winning rate may not be a good thing, because it may also be a prelude to miscalculation. In the first round of the preliminaries, Golaxy's black 113 was the problematic move when losing to Leela Zero (post-match analysis showed that it should have been swapped between A and B to prevent white 114). While white 114 was played, Golaxy's win rate dropped from 49.88% to 33.90%, a range of 16 points—the biggest fluctuation for Golaxy in this game. The preliminaries were played with a 30-minute time limit for each player. In the first round, the timing strategy for the first game was set too fast, and it took us less than 3 minutes to the problematic move 113. We did not have enough time to think. We then adjusted the timing strategy, and the performance became very stable.


Q: I heard that some of the AIs have improved significantly due to the arithmetic support, so has the overall level of this game improved compared to the previous one?

Dr. Jin: The overall level of Go AI has been greatly improved. At the CITIC Securities Cup last August, Korea’s BaduGi team mentioned that they obtained algorithm resources several times higher than before, thanks to the support they received. Before the Fuzhou game, BaduGi told us that they are now much stronger and have a good chance to meet Golaxy in the final. Leela Zero invited Go enthusiasts from all over the world to "run the game" two years ago, and its recent progress is obvious to all. Japan's AQ, which did not participate in this game, also held a press conference recently, announcing that it has support from various aspects such as the Japan Institute of Industrial Technology. The institute has hardware resources of 1088 GPU servers, indicating that Japanese and Korean Go AIs are gaining momentum. Of course, Golaxy has also been improving this year.

Q: Are hardware resources a decisive factor in the level of AI?

Dr. Jin: Resources are one of the most important decisive factors. AlphaGo Zero mentioned in their paper that they used 2000 TPUs and later Alpha Zero used 5000 TPUs. How many resources is that? Let's just do the math. Google is the company with the largest number of servers in the world, TPU is the exclusive hardware developed by Google, which is only available for lease but not for sale. The GPU model VI00 on the market is priced at about 30,000 to 50,000 RMB, with their computing power and price of TPU several times that of GPU. So 2000 TPUs are worth hundreds of millions of RMB, in addition to other auxiliary hardware, electricity, and operation and maintenance fees. It is not difficult for Google to mobilize these resources, but these are not easily acquired and undertaken by small teams. AlphaGo Zero was trained for 40 days, and if we had only one-tenth of their computing power, it might take 400 days to achieve similar results.

The AI computing power and resources of different companies vary greatly. In order to avoid comparison, most of these data will not be disclosed. Golaxy's performance has been greatly improved this year, not because our hardware resources are more powerful, but because it was initially considered that it is difficult for us to compete with large companies in terms of resources And there is also a lack of innovation just from learning existing technologies or copying AlphaGo papers, so we choose to strive to make breakthroughs at the level of algorithms and models.

Q: So is it possible to have an AI competition that aligns computing power?

Dr. Jin: It would be very difficult. You can imagine that even if each AI has the same computational power during the competition, the resources used for normal training are quite different. For example, AlphaGo used more than 200 GPUs in the human vs. AI with Lee Seldo, while AlphaGo used only 4 TPUs in the match with Ke Jie, but 2000 TPUs in the pre-game training. So it's hard to strictly align the computing power.

Q: You and Golaxy have participated in many AI competitions, do you have a preference for the rules and format of the game?

Dr. Jin: Intuitively, there are two forms of AI games. One is a computer-linked game, which at first glance looks like an e-sports game. This type of game has relatively high requirements on the competition platform and requires the organizer to develop and maintain the platform. For example, when I participated in the World AI RYUSEI in Japan in December last year, the network conditions between China and Japan were not good that day, which caused Golaxy to drop off-line in the first round, and was directly sentenced to lose because the event platform did not support reconnection. Later, after negotiation, Golaxy sacrificed part of the time to delay the move to ensure the normal progress of the following rounds. The other form is to play on a physical board and humans will input moves on the board, which is more ritualistic. I can accept both forms.

Q: Are AIs trained in much the same way? How does Golaxy train?

Dr. Jin: To put it simply, AI training mainly involves three things. The first is "game running", which is the most resource-intensive part. The quantity and performance of human games are limited by an upper ceiling, so we use smarter AI to learn the game by continuous practice. The second is training the model, which Leela Zero calls the “weight”. The third is playing, i.e. performing and testing. But in practice, there are many engineering details that need to be taken into account, such as the need for randomness in self-play. Also, a large number of repetitions of playing is not acceptable, so the difference between the Go records requires the adjustment of parameters and configurations. In addition, the structure and features of the model are also a key concern. AI learning is iterative and spiraling. Golaxy plays 10,000-20,000 games per day, each game taking about 1-2 hours, and these games are an important part of Golaxy's reinforcement learning.

After the publication of the AlphaGo paper, we were almost able to reach the same high level, given enough resources and time by the developers. The pioneers put so many resources and effort into making this happen, but there's no need to prove what's already been proven. So I hope to make a further breakthrough in Go AI. Golaxy has on the one hand learned and borrowed from the basic architecture of AlphaGo, and on the other hand has made innovations in feature system, model structure, and MCTS algorithm architecture, including unique area network and score network. Golaxy wants to do things well and do them better. Golaxy uses only a few dozen GPUs of computing power to run the game, and we strive to surpass the highest level of AI Go with fewer computing resources and fewer training samples. In fact, in other fields as well, people always want to use low cost and high efficiency to achieve their goals.
Q: We know that AlphaGo Zero is trained by abandoning the human game records and it was stronger than the Master that used human game records. Does this mean that these game records have no value?

Dr. Jin: Lee, Master, and Zero, three versions of AlphaGo, are getting stronger and stronger in games. The first two versions use the human game records, while Zero discards the human game records and turns to reinforcement learning entirely through self-play. In fact, I think it should be easier to learn based on human games. In this regard, Master has a first-mover advantage over Zero, but its advantage is negligible because the process from zero to the human level is very fast and Zero reached the Lee version in only three days. However, don't overlook AlphaGo's abundant computing resources. Moreover, Zero uses a large network of 40 blocks, which can be interpreted as a larger, more complex model with more brain capacity. If all resources are equal, then learning from human Go records should provide a higher starting point. However, it does not matter—it is the hardware resources and the algorithmic model that matter.

Q: Can we see the ceiling of Go AI now?

Dr. Jin: AlphaGo Zero has such a graph in their paper where the curve for the first 72 hours is rapidly rising, and then the slope slows down. AlphaGo stopped training at 40 days, seemingly hitting a ceiling. I remember a reporter asked at the time why there were only 40 days of training. They answered that the level was strong enough, so they shifted their resources to other programs. If I were in their shoes, I would have decided to keep the program training if I could foresee a significant improvement. The ceiling of Golaxy should be higher than an AI that completely copies the AlphaGo method, at least I think so.

Q: Can you briefly explain to the fans what makes Golaxy different from other AI?

Dr. Jin: First of all, the most distinctive feature of Golaxy is the "non-compromising Go", which continues to expand its winning power when it has an advantage and tries to close the gap when it is disadvantaged. Golaxy's "non-compromising Go" can avoid the problem of an AI’s win rate becoming so low that it makes a messy move.

Secondly, Golaxy can evaluate the progress of the game from two different aspects: win rate and the number of scores, thereby providing a more three-dimensional view of the position. Currently, Golaxy's starting winrate with black is 42%, 1.6 points behind. These two evaluation systems can display data as each move progresses.

Thirdly, Golaxy can set any number of moves and any komi, using both stone count and point count, while most other AIs can only use the Chinese rule of stone counting rule. For example, Golaxy has played in the 17x17 board "Sky Game Go", which starts with four moves in any random position. Then the two players playing black and white start competing (Moyo count), and Golaxy plays 10 komi more, no matter which side the opponent chooses. On March 17th, we partnered with Rui Naiwei to play in the World's Strongest Women's AI Pairs Game in Japan. In the semi-final game, once the game finished, the score was seven points ahead, and Golaxy, playing black, had a 99% chance of winning. If the AI had decided to have 7.5 points komi, it would have judged by losing 0.5 points and there would have been a problem with moves.

Q: In this game, we saw that Golaxy often played "Mi's Flying Dagger joseki". Is this a preferred move?

Dr. Jin: Golaxy didn't play this move much before, but now it loves the move more and more. We have counted the ranking of Golaxy's first 13 moves since the self-play game, and we found that the third move with a san-san, followed by the "Mi's Flying Dagger joseki" is Golaxy's recent favorite, ranking first. My understanding is that if you take the two consecutive stars and then point the corners, then you give White the option to block-edge. Even on the fifth move, a san-san to form the "Mi's Flying Dagger" is the third-ranked option. We have not artificially interfered with avoiding a certain variation, so we can assume that Golaxy has studied this layout very well so far.


Q: Dr. Jin, you seem to be quite good at playing Go.

Dr. Jin: (laughs) A year ago I could play a little bit, less than 10K on FoxGo. Now I'm at 2K or 3K on FoxGo. I think my improvement in Go is not necessarily less than Golaxy's (laughs).

Q: How did you come up with the idea of entering the Go AI field in the first place? Can you tell us about your personal experience?

Dr. Jin: My hometown is Taiyuan, Shanxi Province. I was admitted to Tsinghua University at the age of 14 after winning a national mathematics competition in my senior year. I studied computer science from undergraduate to Ph.D. and then joined IBM, Ali, and Tencent after graduation. In the early days, I worked on cloud computing. Since 2012, I have been engaged in projects related to deep learning, such as image recognition, speech recognition, and advertising precision placement. The work in those years was very fulfilling, and I deeply feel that artificial intelligence will usher in a historic opportunity. This opportunity is mainly reflected in three aspects: one is the increasing abundance of big data; the second is the increasing arithmetic power; the third is the algorithm getting more and more mature. I decided to leave in 2017 and founded Thinker Technology Co., Ltd. in January 2018. When I was working at Tencent, I had a billion user data and hardware resources "regardless of cost". But now we are a small team and we don't have these conditions, so we should work harder on algorithms. I chose to do the Go engine because of my personal interest and also because the data of the Go game is more accessible. That's why I decided to build a technology brand with Golaxy as the starting point of the company.

Q: Has there been anything that stood out to you in developing Golaxy for over a year? What are your plans for future development?

Dr. Jin: The development process has been relatively smooth and the level of Golaxy has been steadily improving. I usually spend most of my time in the company writing programs, debugging codes, holding technical meetings, and going out for games. It may seem boring but it actually was very fulfilling. I was impressed that Golaxy's performance improved significantly after we add our original area network and score network. The entire team is very encouraged to see that the innovative work is producing visible results. At present, there are two versions of Go: one is the high-level game version, and the other is a popular version with weak AI. Now we have developed a mini-program called "Golaxy Coaching", which allows users to play online with the AI Golaxy players at any time, from Golaxy 7D to Golaxy 9D. In my case, my level is about Golaxy 1D or 2D. Golaxy has been developed with a lot of investment so far, and the only financial gain so far is the competition bonus. But unfortunately, it can't be supported by prize money alone, and it needs a commercial project in the future.

At the World Artificial Intelligence Conference in Shanghai last year, Liu He, Vice Premier of the State Council, visited and played with Golaxy on the spot. He asked about the differences between Golaxy and AlphaGo, which shows that China's national leaders not only understand AI but also pay close attention to it. The global artificial intelligence application scenarios and industry scale are growing, which makes me more confident in making the right choice to work in the AI industry.

Q: Thank you Dr. Jin for the interview. I hope the development of Thinker Technology and Golaxy will thrive and prosper.

Dr. Jin: Thank you, and feel free to reach out to Golaxy at any time.


Attachments:
File comment: In March 2019, Jin Xing personally operated Golaxy by teaming up with the legendary Rui Naiwei, and won the championship of the human-AI game.
Picture3.png
Picture3.png [ 456.21 KiB | Viewed 1839 times ]
File comment: Dr. Jin Xing and his AI
Picture2.png
Picture2.png [ 244.11 KiB | Viewed 1839 times ]
File comment: Golaxy (black) vs. Leela Zero
Picture1.png
Picture1.png [ 162.35 KiB | Viewed 1839 times ]


Last edited by 0714952377 on Wed Aug 03, 2022 12:51 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: Interview with Golaxy developer Dr. Jinxing
Post #2 Posted: Wed Aug 03, 2022 12:26 am 
Beginner

Posts: 11
Liked others: 0
Was liked: 2
Rank: Chinese amateur 5 Da
KGS: 1 Dan
Tygem: 3 Dan
IGS: 4 Dan
Wbaduk: 4 Dan
OGS: 7 Dan
Additional photo


Attachments:
File comment: In April 2019, Jin Xing led the Golaxy team on the red carpet at the opening ceremony of the 2nd Wu Qingyuan Cup and World AI Competition.
Picture4.png
Picture4.png [ 473.86 KiB | Viewed 1837 times ]
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 2 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group