Li Shishi why losing AlphaGo? Go player Gu LI: he lightly-chess, artificial intelligence, and Li Shishi-IT information

"The man-machine war of the century" the first inning ended, Li Shishi throw in the towel , which many people have not thought of. The Li Shishi was high hopes, why on earth would lose in the "amateur six or seven" Alphago?

Alphago role of deep learning

"Play chess every step of any of about 35 or so walk, and go go up to 250, multiplied by 250 each step means the whole game appears to almost infinite level. "Google DeepMind lab supervisor Demis-Kazakhstan (Demis Hassabis) said. Alphago in defeating European go champion, Fan Hui studied 30 million game, but after almost six months of study, that number has grown to more than 100 million.

AlphaGo is the core of the depth of the two different neural networks. "Policy networks" (policy network), and "value network" (value network). Their task is cooperation "pick and choose" more promising moves, left obvious poor chess, which will calculate the volume control on your computer can be done in the context of nature and as a human player.

Among them, the "value network" responsible for reducing the depth of the search--AI side of the projection side to judge the situation, the situation obviously disadvantage when you abandon certain routes without a road into black, and "policy network" responsible for reducing the width of the search-in the face in front of a game of chess, some move that obviously should go, such as not sending children to others to eat. Fitting using Monte Carlo, will put the information in a probability function, AI wouldn't have to give every step with the same emphasis, and though they can focus on those moves.

This means that Alphago belongs to the typical reached qifeng, good quick to grab the opponent's weaknesses should be proactive, strike down with strong opponents. Kai-Fu Lee, says the AlphaGo and 1997 compared to the deep blue beat the World Chess Champion, from chess to chess difficult a lot, is very difficult jump.

Deep learning technologies in recent years, a very large amount of data and calculations can be expanded using, than we can imagine. We also known as intelligence, when some false fantasies, in fact, deep learning is growing very fast, it can use more machines very well. So in any objective and scientific assessment of the area, including gaming, finance, search, advertising, and other aspects of the application, humans basically there will be no more chances to compete with machines.

AlphaGo today at the tactical and strategic choice to prove this point, in the case of backward it also has policies, rather than psychological in the psychological warfare. Today AlphaGo errors, then catch up, so we have to think about machine learning and optimization capabilities.

Alphago pre "expression" Li Shishi lightly

Different but Li Shishi, Li Shishi seemed emotional volatility. Li Shishi was born in Korea on a Island, tough and do not throw in the towel, which makes him more down the more brave player. Each player has their own qifeng, today after his dominant instead of errors. Dramatic changes to our surprise this time.

Fan Hui had said that the competition was in early October, 5th to 9th, for five days, is actually two games a day, a total of 10. 5 formal, informal fast chess, there are 5 disks. Official lost, but informal fast chess I won two sets. And Li Shishi today is experiencing a dramatic situation, starting in prevailing circumstances, were eventually reversed up to throw in the towel.

Coulee view, Li Shishi are at a huge advantage in the medium-term, and Alphago are often seen some strange moves, make some silly mistake, which led to Li Shishi of despise your enemy, when the game was over, Li Shishi left about half an hour's time, and Alphago of less than ten minutes left, suggesting that Li Shishi has not spent a lot of time to think, and this is a proof of his enemy.

CSDN Jiang Tao said at the scene, founder, Alphago shown strength fluctuated, lead to false positives in Li Shishi, too lightly, resulting in later got out of hand.

Go first to Ke Jie in thinking machines in the world today needs a breakthrough in psychological, every move is the best, and had no chance of defeating it. Today, it is not the machine needs to break through the psychological, and whether people can break their own.

When two people play chess, you will often observe and ponder each other's emotional and psychological. It is nervous, afraid, you imagine the other side at the same time, this person will also feel the and back. But now the opposite is computer, you face a wall, all your feelings are all being played back, you know it's no fluctuation.

AlphaGo make big mistakes in the early situation, Li Shishi seek win don't want to lose. The stability under the mentality of mistakes, big moves down to later feel, even if the advantage eventually will lose. Li Shishi also feel under a lot of circumstances were unable to bounce back.

Consumption of intelligence, physical exertion, mental stress showed an increasing trend, coupled with public opinion, Alphago "two ears, just better go", but man is the sensor of biological, emotional vulnerable to outside interference.

Ke Jie "program is stronger than human beings it is important that will not be affected by emotional factors. "Level of the machine is probably rushed before professional levels, in spite of still can't get, but infinitely close to professional. Coupled with the machine and lots of learning algorithms, emotional instability, which makes advanced chess player facing the camera is completely different and human opponents in the past.

Today this situation a lot of people didn't expect, as the victory, Li Shishi to lose the match. After this battle, the next four games to look forward to.

李世石为何会输给AlphaGo？围棋国手古力：他轻敌了 - 围棋,人工智能,李世石 - IT资讯

“人机世纪大战”第一局落下帷幕，李世石认输，这是让很多人没有想到的。原本被寄以厚望的李世石，到底为什么会输于古力口中“业余六七段”的Alphago？

Alphago深度学习发挥作用

“国际象棋每步大约会出现35种左右的走位可能，而围棋的走位可能则高达250种，每一步250种相乘就意味着整局比赛会出现多到几乎无穷尽的走位方案。”谷歌DeepMind实验室主管德米斯-哈撒比斯(Demis Hassabis)说道。Alphago在击败欧洲围棋冠军樊麾时学习了3000万盘棋，而经过将近半年的学习，这个数量已经增长到1亿以上。

AlphaGo的核心是两种不同的深度神经网络。“策略网络”（policy network）和“值网络”（value network）。它们的任务在于合作“挑选”出那些比较有前途的棋步，抛弃明显的差棋，从而将计算量控制在计算机可以完成的范围里，本质上和人类棋手所做的一样。

其中，“值网络”负责减少搜索的深度——AI会一边推算一边判断局面，局面明显劣势的时候，就直接抛弃某些路线，不用一条道算到黑；而“策略网络”负责减少搜索的宽度——面对眼前的一盘棋，有些棋步是明显不该走的，比如不该随便送子给别人吃。利用蒙特卡洛拟合，将这些信息放入一个概率函数，AI就不用给每一步以同样的重视程度，而可以重点分析那些有戏的棋着。

这意味着Alphago属于典型的力战型棋风，善于敏锐地抓住对手的弱处主动出击，以强大的力量击垮对手。李开复先生说现在的AlphaGo和1997年击败世界象棋冠军的深蓝相比，从围棋到象棋的难度高了很多，是难度非常大的跳升。

近年来深度学习的技术，非常大的数据量和计算量可以扩张地使用，超过了我们的想象。同时我们也对人所谓的智力，当时有一些错误的幻想，实际上深度学习的成长非常快速，它可以非常好地利用更多地机器。所以在任何客观、科学工程评估的领域，包括游戏，其实是金融、搜索、广告等各方面的应用，人类基本上不会再有更多的机会跟机器来竞争了。

而今天AlphaGo在战术和策略上的选择证明了这一点，在落后的情况下它也有策略，而不是在心理上打心理战。今天AlphaGo先出现失误，随后追赶而上，令我们不得不重新去思考机器的学习和优化能力。

Alphago前期的“表现”让李世石轻敌

但是李世石不同，李世石似乎出现情绪上的波动。李世石出生在韩国一个鸟岛上，坚韧不认输，这让他成为愈挫愈勇的棋手。每个棋手都有自己的棋风，今天他占优势之后反而出现失误。这个时候出现戏剧性变化令我们大跌眼镜。

樊麾曾说比赛是10月初，5号到9号五天，其实是一天两场，一共10盘。5盘正式的，还有5盘非正式的快棋。正式的全输了，但非正式的快棋我赢了两盘。而李世石今天也是经历了戏剧性的局面，在开始占优势的情况下，最后被逆转直至认输。

古力认为，李世石在中期时的优势很大，而Alphago则屡屡出现一些很奇怪的下法，犯一些低级错误，这导致了李世石的轻敌，比赛结束时，李世石还剩半个小时左右的时间，而Alphago只剩不到十分钟，这说明李世石没有花很多时间用来思考，这也是他轻敌的一个证据。

CSDN创始人蒋涛在现场表示，Alphago展现出来的实力忽高忽低，导致李世石出现了误判，过于轻敌，导致后期一发不可收拾。

现今世界围棋第一人柯洁在认为觉得机器在心理方面还需要突破，必须每步棋都是最佳的，人类才完全没有可能战胜它。今天来看，并不是机器需要突破心理，而在与人是否可以突破自己。

两个人下棋的时候，你常常会观察和琢磨对方的情感和心理。它是紧张了，害怕了，你在想象对方的同时，这种作用对方也会感到到，折射回来。但是现在对面是电脑，就是你面对一堵墙，你所有的感觉全部都被打了回来，你知道它没有心态的波动。

在前期AlphaGo有重大失误的局面下，李世石求赢不想输。在求稳的心态下出现失误，心理波动很大，下到后来会觉得，即使优势再大最后也会输。很多情况下李世石自己也已经觉得无力翻盘了。

智力消耗、体力消耗、精神压力呈递增趋势，再加上舆论环境等，Alphago可以“两耳不闻窗外事，一心只下好围棋”，但人类是高感知生物，情绪易受到外界环境的干扰。

柯洁就表示“程序强于人类很重要的一点就是不会受情感因素干扰。”机器的水平大概就是冲职业段之前的水平，虽然职业还到不了，但是无限接近于职业了。再加上机器大量的学习和算法，感情上没有波动，这令高级的棋手面对机器时是和以往面对人类对手完全不同。

今天这个局面很多人都没有想到，胜利在望之际，李世石输掉比赛。经此一役，接下来的四场比赛更值得期待。