Man-machine chess AlphaGo first win the war: Google is a "scheming bitch"-man-machine chess battle, AlphaGo, Li Shishi, go-IT information

The afternoon of March 9 news, man-machine chess showdown began in earnest today, in the game broadcast, dog company CEO Wang xiaochuan was invited to bring users the game analysis and game AlphaGo do the vivid analysis of the principles.

As a staunch supporter of machine learning and technology, Wang xiaochuan analysis is not confined to the AlphaGo for the game itself, but rather look to the on and off the pitch, and stand in the current state of the industry, technology advances and future trend perspectives for netizens to do a multi-dimensional explanation.

Funny thing is, Wang xiaochuan back online analyses why Google is the next big game, the "bitch" in what way? For the first defeat of Li Shishi, Wang xiaochuan, what is the recommendation on machine learning perspective? Recall Wang xiaochuan in the regular analysis of man-machine chess showdown.

Talking about the fight: whore Google

First, Google when looking for Fan Hui, he found a professional chess player at least a State Champion, but the field is relatively low, so that he can get a win, but that is a lot of influence.

There is also a set, Google signed a confidentiality agreement with the Fan Hui, until the cover of Nature before the article was published, Google has not announced have already defeated Fan Hui, beat European champion and is the complement of the article was published.

Finally the Nature cover story published, and immediately played a role in tipping.

So when Google do such a thing, not only in technical research problem, to reflect deeply on what people behind, how the media influence.

I reasoned that in this consideration, Google also made full preparations for this challenge, Li Shishi, is a chess player, in careful thinking about overall things, rather than rashly make the game, later learned that Google′s technology of the action, I think the technology will increase.

About technology: the principle behind AlphaGo

Specifically, AlphaGo has not saved in the memory of one of the games. It of moves are into network parameter has, on seems put all martial arts enrollment operation into a parameter, to you a new of moves zhihou, probably after 13 layer network Diego generation, get a heap number, each place Xia of probability has more big, as long as to he a Board current of layout, he is do of this, each points next of probability has more big on know has.

It very slow when in training, training time, which we know what moves most likely to go under, optimization can do for all the games on the system as a whole, allowing the system to the original collected data as possible to simulate this behavior. So when training is slow, he needs constant adjustment parameters, network, of how such a value, real will become much faster.

It ultimately becomes a function output a 19x19 dimension of such a function is entered is a chess board, output is a function, there is no search or lookup.

Go Greek to me, but as I understand it, people thinking there are several ways, one way is to search, search is calculated, I found him to be so found, I made the following possibilities, which is a way of thinking in the middle.

The second way was my intuition, I could see a lot of chess, the overall shape to local, I will have my own judgment.

The third way or reasoning, walk a few steps to count the breaths. In-depth learning to today, yiqian of computer is search, based of calculation, on reasoning, and intuition is no of, today depth learning has solution has second a problem, can has intuition, in a line-like Xia of when can sentiment how go will better, day of situation whether is optimal of State, put search, and perception this two are master has, now also no reasoning of capacity.

Deep learning is not simply better than the previous game, but all chess had seen the situation, the situation is good, say no to blend into his knowledge of necromancy is unknown. Later he saw the game, whether in whole or in part, he is able to analyze the results.

I did some research, machine methods, and people are like, and locally to find out the difference between human and machine, and algorithm aspects are inseparable. In this case the machine with people in very similar ways of thinking in.

Talk about technology: AlphaGo programming languages in c language

General categories will be divided into machines on principles into three categories, the most basic is the c language, because of its flexibility, the highest calculated up to speed, to build the basic algorithm.

Online, there are two, one is quickly set up some applications, for example, now is the time to do a website, front z has some logic can also make cost-effective, reasonably quick, write programs to quickly find a balance.

There is also a scripting language, that is, each directive itself, the performance will be very slow, but it will be the most convenient, like similar scripting language, I think this thing, this will use the c language as much, you can look up specific information.

Science and technology: why Facebook first go program was robbed by Google?

Facebook did play this thing, but, overall, they also are a team playing on the whole State of mind.

They were early to do this thing, but just some progress post went to places like know how do I do that. Do you think Google is strictly confidential, and the mentality is totally different on both sides.

I wonder if the dog team to do this, you must first establish a powerful aura, a winning aura, but it is regrettable that our engineers have not yet reached this level.

Second, the things that you need a lot of computing resources, Google, use should probably be tens of thousands of the machines node. This declared the 2000 CPU, in which computing resources, has up to 30,000 times times than deep blue computer. The year′s top computer, tell us the milky way, 2nd in the country, one-tenth is the computer, computing power of Tianhe-3 of 300,000, it is 30,000 times, so its computing resources when the need for adequate reserves.

The future: Google AI technology in other areas of application

From AlphaGo the biggest motive, they are a team that is not only used to play chess, can be used to play games, and pharmaceutical research and development, automatic writing program code, he believed that this program is in full swing. The algorithm described in this article is not only a play, its direction is as long as the uncertainty about the two-party game board game, it can also be used, see the Nature of this.

Talk about the threat: status of artificial intelligence and threat

First, artificial intelligence cannot yet complete evolution of the self, now the deep learning method is the data known to repeat, generalizing it, it still has some problems to solve.

Just might have done better, but you cannot create new problems, it also has no creativity, so it does not know the reasoning, and facing new problems.

Artificial intelligence, even if chess people win, but being able to solve the problem is still very limited, just replace repeated mental work.

Worried about artificial intelligence threat to humanity? Now is not the time, the main criterion is:

To see if the machine is conscious. My biggest worry is that military robots, if military robot′s goal is not only to kill to survive, if set, is likely to find his master is an obstacle to his, such as master what power I pulled up, he could have this ability.

If machines are smart enough remains to be seen how it aims to set up, now two, the first machines not purpose, the second not smart enough.

Machine won not terrible, if the machine starts to pretend to lose to you, it′s terrible, he used the other′s intention to hide itself.

Talk about impact: If AlphaGo victory over Li Shishi, what happens next?

If AlphaGo wins the disbanded immediately after playing, because you have to put a team out there, there are people who challenge you, you win meaningless and lose, you die.

So Alphago or Google, this time the strength of greater significance is that it shows, it just wanted to tell you I am cow.

In addition, this competition for in-depth learning and algorithm, will shoulder more responsibility in the entire history in human civilization, the machine will play an increasingly important role on behalf of such power, people should be more on the meaning of their own.

On recommendation: the point of view of artificial intelligence advice to Li Shishi

Google has done to the game a high degree of secrecy, chess style changes are not publicly in the past, so Li Shishi AlphaGo of difficulty is very high. Actually doing so is his mark, the computer changes likely to exceed your expectations.

In turn the wrong place, if he had to understand how computers think it will do, your judgment may be wrong, understanding of computers is not in place, but machine is not good at introducing complex search or do a particular situation, may fall into a pit.

In my judgment on AlphaGo, they will do to Li Shishi no special judgement, he needed a lot of samples to do a game, gets 30 million training game, and Li Shishi game is probably more than 1000, such data is not supported AlphaGo for scale models to support, if he used Li Shishi to training, he will fall into a pit.

Today, the best situation is that both sides have a common heart, do not tell AlphaGo how Li Shishi, improving his skills, Li Shishi also don′t care about AlphaGo routine, try their chess.

Computer speed every second AlphaGo 30,000 times faster than ever before, and at that rate, continuous training, developed about two beats each other, keep yourself practicing.

When losing yourself lessons as soon as this place lost, holding games what to do to improve, his own teacher, such an approach can improve very quickly, is also in a closed board game machine quickly defeated the essence of people.

人机围棋大战AlphaGo获首胜：谷歌是个“心机婊” - 人机围棋大战,AlphaGo,李世石,围棋 - IT资讯

3月9日下午消息，人机围棋对决今日正式打响，在比赛直播中，搜狗公司CEO王小川应邀为网友带来了比赛分析，并就比赛中AlphaGo做了技术原理方面的生动分析。

作为机器学习和技术的坚定支持者，王小川对比赛的分析没有局限于AlphaGo本身，而是放眼到了赛场内外，并站在行业当前状态、技术进展和未来趋势等角度为网友做了多维度讲解。

有趣的是，王小川还给网友分析了为什么这次谷歌是在下一盘很大的棋，其“心机婊”表现在哪些方面？对于首场落败的李世石，王小川又有什么关于机器学习角度的建议？一起来回顾一下王小川在人机围棋对决中的经常分析。

谈比赛：心机婊谷歌

第一，谷歌找樊辉的时候，他找了一个职业棋手至少是一个洲的冠军，但其实段位相对比较低，这样他能够在里面得到一个取胜，但是说出去是很大的影响力。

另外还有一个设定，谷歌和樊麾签了保密协议，直到《Nature》封面文章发表前，谷歌都没有对外宣布已经战胜了樊麾，战胜欧洲冠军和文章发表是相互配合的。

最后等到《Nature》封面文章一发表，立刻起到了引爆的作用。

所以谷歌在做这件事情的时候，不仅在做技术研究的问题，还深刻考虑了背后选什么人，怎样实现媒体影响力。

我推断在这种考虑下，谷歌对于这次挑战李世石也做了充分的准备，是一个围棋选手在缜密的思考全盘的事情，而不是莽撞地做这个比赛，后来也了解到了谷歌在技术之外的整个动作，我认为它的技术会增大。

谈技术一：AlphaGo背后原理

具体来说，AlphaGo的内存里已经不用存一个一个的棋谱了。它的棋谱都变成网络参数了，就好像把所有武功招术变成一套参数，给你一个新的棋谱之后，大概经过13层网络迭代，得到一堆数，每个地方下的概率有多大，只要给他一个棋盘当前的布局，他正在做的这个，每个点下一步的概率有多大就知道了。

它在训练的时候很慢，训练的时候，我们知道当下棋谱下最有可能走哪个走，系统整体上能够为所有的棋谱做优化，使系统在原有的采集到的数据上尽可能地模仿这样一个行为。所以训练的时候是慢的，他需要不断的调整参数，需要怎么样的网络叠加出这样一个值来，实际下的时候会变得快很多。

它其实最终是变成一个函数，输出一个19×19维的这样一个函数，就是输入是一个棋盘，输出是函数，过程中没有搜索或查找过程。

我对围棋不懂，但我的理解是，人有几种思考的方式，一种方式是搜索，搜索就是计算的时候我这么搜他这么搜，我把下面的可能性展开，这是中间的一种思考的方式。

第二种方式就是我的一种直觉，我可能看多很多棋，对整体对局部的行状，我会有我的判断。

第三种方式是推理，走几步去数一下几口气。深入学习到今天，以前的计算机是搜索，基础的计算，对推理、直觉是没有的，今天深度学习已经解决了第二个问题，能够有直觉，在一个行状下的时候可以感悟怎么走会更好，当天的局势是否是最优的状态，把搜索、知觉这两个都掌握了，现在还没有推理的能力。

深度学习里不是简单地比对以前的棋，而是把以前见过的所有棋的局面，这个局面好不好，都融汇贯通变成他说不清道不明的知识。以后他看到这个棋，不管是整体还是局部的，他都能分析出一个结果来。

我之前做过研究，机器做的方法和人是很像的，局部找出人和机器的不同点，算法层面上是分不开的。这种情况下机器跟人用了很类似的思考方式在进行。

谈技术二：AlphaGo的程序语言可能以C语言为主

一般大类上会分成机器原理分成三个类别，一种最基础是C语言，因为其灵活性最高，计算起来也能够快，也能构建最基层算法。

网上有两种，一种是用快速搭建一些应用，比如说我们现在要去做一个网站的时候，前面用Z既有一些逻辑也可以做成性价比，在性能合理的快速里面，写程序快速里面找到一个平衡。

还有一个就是脚本语言，就是每个指令本身的性能会很慢，但是写起来会最方便，就像类似的脚本语言，我认为这个事，这个会以C语言为多，可以再查一下具体的资料。

谈科技：先做围棋程序的Facebook为何让谷歌抢了先？

Facebook也在做下棋这个事情，但是整体来看，他们整体上还是一个团队在玩的心态。

他们是更早做这个事情的，但是随便有一些进展就跑到类似知乎一样的地方发帖讲我是怎么做的。你看谷歌是严格的保密，双方的心态是完全不一样的。

我在想如果搜狗的团队去做这个事，就一定要先建立起一个强大的气场，一个必胜的气场，但遗憾的是我们的工程师还没有达到这样的高度。

第二，这个事情你需要大量的计算资源，谷歌在这次里面，大概动用了应该是上万台的机器节点。这次对外宣称有2000个CPU，在这样计算资源的时候，已经比当年深蓝计算机提升了3万倍。也是当年顶级的计算机，咱们讲的这个国内的天河2号，是它的计算机的十分之一，天河计算力的3的30万，它是3万倍，所以它的这样计算资源的时候也是需要有足够的储备。

谈未来：谷歌AI技术的其他领域应用

从AlphaGo最大的目的，他们的团队认为不仅是用来下棋，可以用来玩游戏、医药研发、自动写程序写代码，他认为这套程序是可以全面铺开的。这篇文章中讲到的算法不仅是下围棋，它的方向是只要是对两方博弈的不确定性的棋盘游戏，它都能够同样适用，《Nature》看到了这一点。

谈威胁：人工智能的现状和威胁

第一，人工智能现在还不能完成自我的进化，现在深度学习的方法只是对已知的数据去重复，泛化它，其实解决的还是已经有的问题。

只是可能做得更好，但不能创造新的问题，它还没有创造力，所以它不会懂得推理和面对新的问题。

人工智能，即便下棋把人赢了，但能够解决的问题还是非常有限的，只是取代重复性的脑力劳动。

担心人工智能对人类有威胁？现在还不到时候，主要判断标准是：

要看机器是否有意识。我最大的担心是军用机器人，如果让军用机器人的目标不仅是杀敌还要存在下去，如果设置存在下去，有可能发现他的主人是他的障碍，比如主人会怎么把我的电源拔了，他可能会产生这种能力。

如果机器足够聪明的话还要看它的目的是怎么设定的，现在两个，第一机器没有目的，第二不够聪明。

机器赢了不可怕，如果机器开始假装输给你，那就可怕了，他用其他的意图来隐藏自己。

谈影响：如果AlphaGo接下来完胜李世石，接下来会怎么样？

如果AlphaGo赢了之后就会立刻解散不玩了，因为你要放一个团队在那儿，总有人挑战你，你赢了没有意义，输了就挂掉了。

所以说Alphago也好，谷歌也好，这一次更大的意义在于它的实力展示，它就是想告诉大家我很牛。

另外，这次比赛对于做深度学习和做算法，在整个历史里会扛起更多的责任，在人类文明里，机器会扮演越来越重要的作用，代表这样的力量，人应该更多思考自己的意义。

谈建议：以人工智能的角度对李世石的建议

谷歌对这次比赛做了高度的保密，以往下棋风格怎么变化都没有公开，所以李世石学习AlphaGo的难度非常高。实际上这么做也是刻舟求剑，计算机的变化有可能超出你的预期。

反过来错误的地方，如果他去理解计算机怎么想，机器会怎么做，有可能你的判断是错的，对计算机的理解不到位，反而机器不擅长引入复杂局或者做某种特定情况的搜索，可能会掉到一个陷阱里去。

以我对AlphaGo的判断，他们并不会对李世石做特别的判断，他需要大量的棋局样本去做，要拿三千万的棋局做训练，拿到李世石的棋局可能就是一千多盘，这样的数据规模是远远没法支撑AlphaGo的模型去支撑，如果他用李世石做训练，他也会掉到一个陷阱里去。

今天最好的状况是双方都有一颗平常心，AlphaGo也别讲李世石怎么样，就是提高他的棋艺，李世石也不要管AlphaGo的套路，尽力去下自己的棋。

电脑每秒钟的速度比以前AlphaGo的速度快了三万倍，以这个速度不断自我训练，同时开发两个脑左右互搏，不断地自己跟自己练。

自己跟自己下输的时候就马上总结教训，这个地方输掉了，拿着棋局怎么去做改进，自己当自己的老师，这样的方法能提升得非常快，也是在封闭的棋类游戏中机器迅速战胜人的精髓所在。