An AI program developed by Google DeepMind that has mastered the ancient game of Go won this year’s Innovation Lions Grand Prix in Cannes.
Go is a game invented in China over 5500 years ago, formerly regarded as one of the four cultivated arts of the scholar gentleman. It is considerably more complex than chess, possessing more possible moves than the total number of atoms in the visible universe, which has made it so far impossible for any computer to master. AlphaGo’s genius lies in its ability to play using something like intuition, just as a human player would.
阿尔法狗，由谷歌DeepMind研发｜AlphaGo, developed by Google DeepMind
创新狮子评审会主席Emad Tahtouh总结说道，“从任何角度衡量 — 不论是复杂性，简洁性，或者创新性，它的潜在应用或者是现在的用途，还有它的成功之处- 阿尔法狗都令人无法置信。它包含了我们要寻找的创新的所有方面。在全世界很多其他领域里，其潜力都巨大无穷。”
Innovation Lions jury president Emad Tahtouh summarised, “AlphaGo, by any measure – whether you’re looking at complexity or simplicity or innovation or its potential use, its current use, its success – is incredible. It encapsulates everything we’re looking for in innovation. Its potential throughout the world and in so many other avenues is incredible.”
阿尔法狗登上《自然》杂志首页｜AlphaGo on front cover of Nature magazine.
AlphaGo’s capabilities were most clearly demonstrated when it defeated one of the worlds greatest players, Lee Sedol, in a five-game match broadcast live in March this year, for a $1 million prize. Lee Sedol lost the series 1-4, though it is worth noting that he won one of the matches by pulling off a completely unexpected and seemingly illogical move, which outfoxed AlphaGo and has since been described as “a masterpiece [that] will almost certainly become a famous game in the history of Go”.
3月10日韩国首尔，阿尔法狗对弈李圣石，全球顶尖围棋选手｜Seoul, South Korea, March 10th. AlphaGo vs. Lee Sedol, one of the greatest players in the world.
Google DeepMind is a subsidiary of Google that researches into artificial intelligence. Following the Innovation Lions pitch session at Cannes Lions, SHP+ caught up with Thore Graepel, Research Lead at Google DeepMind and Professor of Computer Science at University College London (UCL) to go a little deeper.
Thore Graepel：1996年Deep Blue在象棋比赛上赢了Garry Kasparov之后，我们就想着要去对付更为复杂得多的围棋，挑战巨大。而当时，计算机很不擅长围棋，因为围棋极其复杂，人类在下棋时会运用自己的感知还有直觉和创造力。而我们的挑战就是能否复制人类的智慧？能不能开发出计算机程序，可以将人类在下棋时的逻辑推理和直觉力结合起来，只去看可以让你获胜的步数，来正确判断整体格局。所以我们利用了神经网络。
SHP+: What was the genesis of the AlphaGo project?
Thore Graepel: Ever since Deep Blue beat Garry Kasparov at chess (1996) it was this big remaining grand challenge to tackle a much more complex game. Computers at that point were very poor at [Go, because it] is incredibly complex and humans apply a lot of perception and intuition and creativity in the face of that. The challenge for us was, can we replicate that? Can we create a computer program that combines the reasoning you use when you play a game with the intuition of really only looking at promising moves and judging the overall position correctly. We then used a neural network approach to tackle that.
1996年，“深蓝”击败世界象棋大师Garry Kasparov｜1996, Deep Blue defeated Garry Kasparov, world chess Grandmaster
SHP+: What was the timeline of the project?
Thore Graepel: The project started out as an internship project. Once it was clear that these neural-networks were very good at playing Go, we doubled down on the project and added many more people. It was a two-year project; the first was this early internship phase and the second year we had anywhere between ten and twenty people working on it.
李圣石，职业围棋九段选手｜Lee Sedol, professional Go player of 9 dan rank
SHP+: When did you first test AlphaGo against a human player?
Thore Graepel: We needed a human evaluation…to show that we could actually beat a professional Go player. We asked Fan Hui (Chinese-born French, three time Go European champion) and AlphaGo beat him 5-0…which was great news for us because we had only evaluated AlphaGo against other Go programs up until that point. Fan Hui was a bit devastated, but he came around and grasped that this is something great that the game of Go can really benefit from.
Thore Graepel：人们的反映还是挺复杂的。一开始很多人没料到阿尔法狗能打败李世石。但是当阿尔法狗真的赢了一局的时候，这些人不得不去接受这样的现实。之后，比赛后期，人们变得很狂热。并且还给阿尔法狗颁发了九段证书（最高级别），人们也真正开始接受这样的想法，知道他们可以从阿尔法狗身上学到一些东西。一如当初Deep Blue并没有对象棋产生任何消极影响一样，我觉得阿尔法狗的存在也不会对围棋产生任何负面影响。
SHP+: After his match with AlphaGo, Lee Sedol said, “As a professional Go player, I never want to play this kind of match again.” How does the worldwide Go community feel about AlphaGo?
Thore Graepel: They have mixed feelings about it. A lot of people at the beginning didn’t think that AlphaGo would be able to beat Lee Sedol. Then, when it started winning, they had to adapt to the idea that it might be possible. Then, toward the end, people were so enthusiastic about it. They gave Alpha Go the 9 dan (top level) pro certificate and they really started to embrace the idea [and] that they might learn a lot from Alpha Go. Just as Deep Blue hasn’t had any negative effect on chess, I think the existence of Alpha Go has had no bad effect on Go.
一台雅达利2600家用游戏机｜An Atari 2600 Console
Thore Graepel: 这要追溯到《自然》杂志有关DeepMind的第一次报道，记录了从雅达利2600家用游戏机中找出的50个街机游戏，并且培训了一个单一特工，单一的神经网络，去打每一个游戏……每一个游戏都像是一个小宇宙一样。比如说“吃豆人”，有迷宫，有空间导航，有人追你，你要逃跑，得吃豆子。这是一个非常丰富的小世界。显然，人工智能学会了如何打游戏，这就意味着它对这个游戏世界有了一定的了解。而现在是50个这样的世界，所有这些游戏加在一起，涵盖了认知的各个不同的方面，这样的测试非常有意思。
SHP+: Why are games a good tool for developing AI?
Thore Graepel：It goes back to the first Nature (science journal) paper that came out with DeepMind which was the idea of taking 50 arcade games from the Atari 2600 console and training a single agent, a single neural network, to be able to play each one of these games…Every single game is…like a microcosm of the world. Think of Pacman; there’s a maze, there’s spatial navigation, things pursue you, you have to escape, you have to eat pills. It’s a very rich little world. Clearly AI learning to play that game would mean it had understood something about the world. Now take that times 50, where all of these games cover different aspects of cognition and you have a really interesting test.
That’s the foundation of what we do. We look for these games…that capture an aspect of the real world and then we train AI agents to cope with those. Or if they can’t, then we invent something new.
《吃豆人》游戏｜The game Pacman
SHP+: What can be done with tech in future?
Thore Graepel: The beautiful thing is that the way we solved the problem is very general, it combines these ideas of learning and reasoning. So for example, you can think of applications in healthcare or research where people who can be assisted to make the right choices, where the AI can read much more of the medical research literature, and therefore give better hints as to what the next step is, and maybe plan better ahead, to assist researchers or doctors.