American Go E-Journal » Master/AlphaGo Commentaries

AlphaGo is retiring. DeepMind’s Demis Hassabis and David Silver made the stunning announcement as the Future of Go Summit wrapped up in Wuzhen, China, saying that the match against world #1 Ke Jie represented “the highest possible pinnacle for AlphaGo as a competitive program” and would be the AI program’s final match.

“The research team behind AlphaGo will now throw their considerable energy into the next set of grand challenges, developing advanced general algorithms that could one day help scientists as they tackle some of our most complex problems, such as finding new cures for diseases, dramatically reducing energy consumption, or inventing revolutionary new materials,” Hassabis said. “If AI systems prove they are able to unearth significant new knowledge and strategies in these domains too, the breakthroughs could be truly remarkable. We can’t wait to see what comes next.”

DeepMind isn’t leaving the go community empty-handed, however. As a “special gift to fans of Go around the world,” DeepMind is publishing a special set of 50 AlphaGo vs AlphaGo games, which Hassabis and Silver said “we believe contain many new and interesting ideas and strategies for the Go community to explore.”

And while DeepMind doesn’t plan to give AlphaGo itself a wide release, Hassabis says he’s more than happy for others to make use of DeepMind’s research themselves. Programs like Tencent’s Fine Art and Japan’s DeepZenGo have used similar deep-learning techniques to achieve around 9th-dan level, according to Hassabis. DeepMind will soon publish another paper on how it architected the latest version of AlphaGo, AlphaGo Master, and Hassabis expects other companies to learn from the new research.

Also, Hassabis said that “We’re also working on a teaching tool – one of the top requests we’ve received throughout this week. The tool will show AlphaGo’s analysis of Go positions, providing an insight into how the program thinks, and hopefully giving all players and fans the opportunity to see the game through the lens of AlphaGo. We’re particularly honoured that our first collaborator in this effort will be the great Ke Jie, who has agreed to work with us on a study of his match with AlphaGo. We’re excited to hear his insights into these amazing games, and to have the chance to share some of AlphaGo’s own analysis too.”

Read more in The Verge and on the DeepMind website. photo courtesy The Verge

AlphaGo completed its sweep of world number one professional Ke Jie 9P on Saturday, winning the third and final game of their match by resignation. Ke called it “one of the greatest matches that I’ve had.” The game once again showcased exciting and surprising moves from both sides, the first arising almost immediately on move 7, a four-space extension from the upper right in which AlphaGo played one space closer to the corner than in the usual Chinese opening. White 20 was a counter-intuitive second-line probe into Black’s framework on the lower right, showcasing Ke Jie’s superb positional judgment.

When Ke Jie attained a local advantage in the centre, AlphaGo switched to build a powerful framework on the top that spurred White to invade. The action came to a head when Ke Jie sacrificed the territory on the upper side to AlphaGo, gaining initiative to pressure the lower left. After AlphaGo protected its group, the match proceeded towards the endgame. Ke Jie revived his stones in the upper left to take the territorial lead, but this sequence left AlphaGo just enough latitude to take control of his group in the centre, and White resigned after 209 moves.

“We held this event aiming to discover new insights into this ancient, beautiful game,” said DeepMind CEO Demis Hassabis. “I can safely say that what has taken place since Tuesday has exceeded our highest hopes. We have seen many new and exciting moves, and we also saw AlphaGo truly pushed to its limits by the great genius Ke Jie.”

Adapted from a report on DeepMind’s AlphaGo page.

“Playing games like this will give us new ideas about how to play,” said Gu Li 9P, after playing in the AlphaGo-Pair Go and commentating on the Team Go event. “It felt like four painters working together on a shared canvas,” added AlphaGo Lead Researcher David Silver, “all with different styles, all combining together to make something truly beautiful.”

In Pair Go, the first of the day’s matches on Thursday, top Chinese professionals Gu Li and Lian Xiao each had their own AlphaGo teammate, alternating moves in tag team style. In the second, Team Go, five of China’s top professional Go players had the unique challenge of working together to take on AlphaGo’s distinctive style.

In Pair Go, AlphaGo and its professional teammate agreed with each others’ moves – though they surprised each other from time to time too. In a sense, the match provided a glimpse of how human experts might be able to use AI tools in the future, benefiting from the program’s insights while also relying on their own intuition. The AlphaGo/Lian Xiao Pair Go team prevailed over AlphaGo/Gu Li, winning by resignation.

Team Go provided a different but no less compelling challenge, requiring players to coordinate closely to make the most of the format. The professional teammates – Zhou Ruiyang, Chen Yaoye, Mi Yuting, Shi Yue and Tang Weixing – had access to their own study board to discuss and analyse variations, allowing them to draw on centuries of Go wisdom and styles as they debated strategies. They approached the challenge in a light-hearted manner, clearly enjoying the experience of playing together, and their resulting style was very balanced. In the end, AlphaGo, once again, won by resignation.

“AlphaGo could actually broaden the horizon of Go playing,” said Lian Xiao. “It could bring more imagination into Go.”

The final game between AlphaGo and Ke Jie will be played at 10:30p EDT Friday night; DeepMind is streaming the matches live, posting match updates and expert commentaries every day on this page and on their Twitter account, @DeepMindAI. For more details, you can visit the official event page here

- adapted from a report on the DeepMind AlphaGo website

Despite 100 moves that “were the best anyone’s ever played against the Master version,” world number 1 Ke Jie 9P was forced to resign Game 2 of his match against AlphaGo on Thursday in Wuzhen, China, clinching the best-of-3 series for the go AI. Afterwards, Ke said that he thought he “was very close to winning the match in the middle of the game” and that he was so excited “I could feel my heart thumping!” But, he admitted, “Maybe because I was too excited I made some stupid moves. Maybe that’s the weakest part of human beings.” The latest version of AlphaGo, Ke added, “is 100 percent perfection…For human beings, our understanding of this game is only very limited.”

The game was extraordinarily complex, with seven separate groups on the left and lower sides, all of them interrelated and none of them settled. This type of complex interaction, impossible to calculate fully and demanding the most of each player’s value judgment and intuition, brought both Ke Jie and AlphaGo into their element.

With many groups hanging in the balance, both sides continued raising the stakes. Ke Jie played daringly, creating the possibility of sacrificing the ko and two of his groups to take AlphaGo’s two groups in the upper left on an even larger scale. However, AlphaGo chose to settle the ko and the game by connecting at move 137, conceding enormous gains to White on the lower left to secure even greater profits in the lower right. As Ke Jie, playing white, could not control the whole upper left, AlphaGo’s territorial advantage proved decisive.

“What an honor it is to play with a genius like Ke Jie,” said Demis Hassabis, CEO and co-founder of DeepMind. “This is called the Future of Go Summit, and today I think we saw a game from the future,”

Still to come are Pair and Team Go on Friday, and the third AlphaGo-Ke Jie match on Saturday. (use this Time Zone Converter to determine local dates/times)

DeepMind is streaming the matches live, posting match updates and expert commentaries every day on this page and on their Twitter account, @DeepMindAI. For more details, you can visit the official event page here. American Go Association chapters continue to play watch parties (they’re eligible for $100 in non-alcohol expenses like pizza; click here for details); email details to journal@usgo.org and we’ll post an updated report.

- adapted from a report on the DeepMind/AlphaGo site; photo by China Stringer Network, via Reuters

Other match coverage:
Google’s A.I. Program Rattles Chinese Go Master as It Wins Match (New York Times)
AlphaGo beats Ke Jie again to wrap up three-part match (Verge)
Google’s AlphaGo Continues Dominance With Second Win in China (Wired)
China censored Google’s AlphaGo match against world’s best Go player (The Guardian)

by Andy Okun, reporting from the  ‘Future of Go’ summit in Wuzhen, China

The version of AlphaGo that defeated Ke Jie 9p in the first round of the three game challenge match yesterday was trained entirely on the self-play games of previous versions of AlphaGo, a Google DeepMind engineer told an audience in China.  David Silver (at right), lead researcher on the AlphaGo project, told the Future of AI Forum in Wuzhen that because AlphaGo had become so strong, its own games constituted the best available data to use.

The version of AlphaGo that beat Fan Hui 2p in 2015 (AlphaGo Fan) and the one that defeated Lee Sedol 9p last year in Seoul (AlphaGo Lee) each included a “value network,” designed to evaluate a position and give the probability of winning, and a “policy network,” designed to suggest the best next move, that were trained using hundreds of thousands of skilled human games.  The most recent version, AlphaGo Master, trained both networks on a database of its self-play games generated by its predecessors.

This was not the only new information Silver revealed about system.  The version playing Ke Jie is so much more efficient that it uses one tenth the quantity of computation that Alphago Lee used, and runs on a single machine on Google’s cloud, powered by one tensor processing unit (TPU).  AlphaGo Lee would probe 50 moves deep and study 100,000 moves per second.  While that sounds like a lot, by comparison, the tree search powering the Deep Blue chess system that defeated Gary Kasparov in the 1990s looked at 100 million moves per second.

“AlphaGo is actually thinking much more smartly than Deep Blue,” Silver said.

In addition, Silver revealed that DeepMind had measured the handicap needed between different versions of the software. AlphaGo Fan could give four stones to the previous best software, such as Zen or CrazyStone, which had reached 6d in strength. AlphaGo Lee, in turn, could give AlphaGo Fan three stones, and AlphaGo Master, which at the new year achieved a 60-game undefeated streak against top pros before coming to this challenge, is three stones stronger than AlphaGo Lee.  Silver delivered this with the caveat that these handicap stones are not necessarily directly convertible to human handicaps.  Professional players suggested that this may be due to AlphaGo’s tendency to play slowly when ahead — i.e., an AlphaGo receiving a three stone handicap may give its opponent ample opportunities to catch up, just as yesterday’s AlphaGo let Ke Jie get to a 0.5 point margin. This also reveals that AlphaGo is able to play with a handicap, previously a matter of speculation in the go community.

Silver’s talk came after DeepMind chief Demis Hassabis gave a passionate account of how go and AI research have fed each other. Go is so combinatorially large that playing it well is intuitive as well as a matter of calculation.  The methods that have worked so well with AlphaGO have generated moves and strategies that seem high level, intuitive, even creative. These same methods have applications in medicine, energy and many other areas. He quoted Kasparov: “Deep Blue was the end.  AlphaGo is the beginning.”

photos by Dan Maas