【视角呆萌】阿尔法汪的优良与缺陷?人工智能展望与投资(三)
<span style="color: black;">★</span><p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;">前言:2016年3月13日,李世石在连续3场被阿尔法汪血虐后重整旗鼓,在第4场白78步下出“神之一手”,之后阿尔法汪连续懵逼失误连连,<span style="color: black;">最后</span>小李上演王者归来。<span style="color: black;">咱们</span>在系列一中以算法为核心的分析似乎得到印证,本文试图再次从算法的<span style="color: black;">规律</span>出发,分析阿尔法汪的<span style="color: black;">优良</span>与缺陷,让<span style="color: black;">咱们</span>更加<span style="color: black;">安然</span>地面对日益到来的人工智能。</span></strong></span></p><span style="color: black;">★</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><img src="http://mmbiz.qpic.cn/mmbiz/LaSSr9wZrk5T3Aprhxib6E1usJnK87icnc0GNgyMrIefOnJLfibNAow7NicAPicDBWEU37MJXMCsMQDrVHTW72p4T5A/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;"></strong></span><strong style="color: blue;">“神之一手”与阿尔法汪的懵逼</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><img src="http://mmbiz.qpic.cn/mmbiz/LaSSr9wZrk5T3Aprhxib6E1usJnK87icncV176yrPbdPibkTsYytETLeiaiaZrLyQly80qQXZsUXCBBInCACiba9laVQ/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;"></strong></span><strong style="color: blue;"><span style="color: black;"> 第四局比赛,阿尔法汪前期下出了一个“囧”字,在开局<span style="color: black;">把握</span>了局势。李世石在大势落后的<span style="color: black;">状况</span>下,下出绝妙的白78手,被古力等解说盛赞为“神之一手”。此后阿尔法汪似乎忽然陷入BUG,初学者水平的无用手连续<span style="color: black;">显现</span>,<span style="color: black;">最后</span><span style="color: black;">长期</span>计算后落败。</span></strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;"><img src="http://mmbiz.qpic.cn/mmbiz/LaSSr9wZrk5T3Aprhxib6E1usJnK87icncospPvibSAiaBIokmsUVxuKGkfcXXeToAX2aRHxZYarHsz4kBoqMaaxRQ/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">Deep Mind创始人Demis Hassabis在Twitter中<span style="color: black;">暗示</span></span><strong style="color: blue;">阿尔法汪的估值网络(用来评定局面胜率的算法)在79步后<span style="color: black;">显现</span>了严重错误。它<span style="color: black;">始终</span>判断自己<span style="color: black;">处在</span>70%胜率的<span style="color: black;">优良</span>状态,<span style="color: black;">因此</span>以快速<span style="color: black;">处理</span>战斗的方式<span style="color: black;">起始</span>乱下(<span style="color: black;">她们</span>的算法<span style="color: black;">便是</span>赢就OK,不在乎赢多少目),<span style="color: black;">而后</span>在87步后<span style="color: black;">发掘</span>对局势判断错误。</strong><span style="color: black;">李世石在赛后采访<span style="color: black;">亦</span><span style="color: black;">暗示</span>“我觉得AlphaGo并不完美,肯定有弱点的,感觉大致有两点,他执黑下得并不太好,<span style="color: black;">另一</span>当我下出意外一手,AlphaGo应对可能就会<span style="color: black;">显现</span>失误。</span><strong style="color: blue;">当我下出完全<span style="color: black;">无</span>想到的棋,AlphaGo<span style="color: black;">全部</span>程序似乎就会<span style="color: black;">显现</span>问题。</strong><span style="color: black;">”</span></p>
<span style="color: black;"><strong style="color: blue;">阿尔法汪的<span style="color: black;">优良</span>与缺陷</strong></span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"> 部分内容引用陈经(香港科技大学计算机科学硕士,中国科学技术大学科技与战略风云学会<span style="color: black;">科研</span>员)</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">《<span style="color: black;">设备</span>完胜后分析AlphaGo算法巨大的<span style="color: black;">优良</span>与可能的缺陷》</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">咱们</span>在系列(一)中讨论过,阿尔法汪<span style="color: black;">是由于</span>4个系统<span style="color: black;">构成</span>,并用通俗易懂的语言给予解释,<span style="color: black;">这儿</span>再回顾一遍:</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">走棋网络(Policy Network),给定当前局面,预测/采样下一步的走棋。</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">(给每种下棋的可能步打个分)</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">快速走子(Fast rollout),<span style="color: black;">目的</span>和1<span style="color: black;">同样</span>,但在适当牺牲走棋质量的<span style="color: black;">要求</span>下,速度要比1快1000倍</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">(打分不是<span style="color: black;">那样</span>准,<span style="color: black;">然则</span>打的快)</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">估值网络(Value Network),给定当前局面,估计是白胜还是黑胜。</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">(<span style="color: black;">按照</span>阿尔法汪看的3000万局棋,判断盘面胜率,讲道理<span style="color: black;">倘若</span>我能看3000万场我<span style="color: black;">亦</span><span style="color: black;">晓得</span>谁赢啊。。)</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">蒙特卡罗树搜索(Monte Carlo Tree Search,MCTS),把以上这三个部分连起来,形成一个完整的系统。</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">(把<span style="color: black;">有些</span>一看就很笨的着数去掉,<span style="color: black;">而后</span>自己脑补后面下棋<span style="color: black;">各样</span>可能,看结果后再<span style="color: black;">选取</span>。其实咱们<span style="color: black;">亦</span><span style="color: black;">这般</span>,只是想多了脑袋大呀)</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">优良</span>一:“细腻”的大局观与<span style="color: black;">持续</span>训练的“打分系统”。</span></strong><span style="color: black;">咱们人类的大局观<span style="color: black;">一般</span>会<span style="color: black;">经过</span>经验与灵感去思考,而阿尔法汪则会<span style="color: black;">连续</span>在历史经验的<span style="color: black;">基本</span>上去细细考究每条胜率,<span style="color: black;">况且</span><span style="color: black;">经过</span><span style="color: black;">一直</span>的“复盘”<span style="color: black;">连续</span>去验证。</span><span style="color: black;">Z</span><span style="color: black;">en和CrazyStone等上一代程序,以及facebook田渊栋博士<span style="color: black;">研发</span>的Darkforest都用了MCTS。它们和AlphaGo虽然棋力相差很远,<span style="color: black;">然则</span>行棋思想其实很<span style="color: black;">类似</span>。 </span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="http://mmbiz.qpic.cn/mmbiz/LaSSr9wZrk5T3Aprhxib6E1usJnK87icnc4GzHbf4T3ibBbBRUfOAsXTbNzY9Yl9EUuMXqHe77zNYGqHibOnYvibic2A/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;"><img src="http://mmbiz.qpic.cn/mmbiz/LaSSr9wZrk5T3Aprhxib6E1usJnK87icnceKoIlibtAycnJO9j6WGwDsydfN0K9TCtEqQpmFqtw1ERHjUBJDDI0gg/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"> 这是田渊栋贴的Darkforest对前两局的局势评分。</span><strong style="color: blue;"><span style="color: black;">能够</span>看出,这个评分和棋局走向高度一致,完全说得通。<span style="color: black;">况且</span>谷歌<span style="color: black;">亦</span>透露了AlphaGo对局势的评分,虽然<span style="color: black;">始终</span>领先,但第二局<span style="color: black;">亦</span>有接近的时候,能够相互印证。</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">(技术讨论<span style="color: black;">能够</span>看<span style="color: black;">这儿</span>哦:</span></strong><span style="color: black;">所说</span>的局势评分,<span style="color: black;">便是</span>程序的MCTS模块,对模拟的<span style="color: black;">恰当</span>局面的胜率估计,都是从当前局面,<span style="color: black;">选取</span><span style="color: black;">有些</span>分支节点搜索,<span style="color: black;">始终</span>分支下去到某层的“叶子”节点,<span style="color: black;">例如</span>深入20步。<span style="color: black;">这个分支策略,AlphaGo和Darkforest用的是“策略网络”<span style="color: black;">供给</span>的选点,选概率大的先试,又鼓励没试过的走走。到了叶子节点后,就改用一个“快速走子策略”<span style="color: black;">始终</span>下完,不分支了,你一步我一步往下推进,<span style="color: black;">例如</span>再下200步下完数子定出胜负。这个走子策略<span style="color: black;">必要</span>是快速的,谷歌论文中说AlphaGo的快速走子策略比策略网络快1000倍。<span style="color: black;">倘若</span>用策略网络来走子,那就<span style="color: black;">无</span>时间下完了,和李世石对局时的2小时会远远<span style="color: black;">不足</span>用。下完以后,将结果一路返回,作<span style="color: black;">有些</span>标记。最后统计所有<span style="color: black;">恰当</span>的<span style="color: black;">最后</span>局面,看双方胜利的各占多少,就有一个胜率报出来,<span style="color: black;">做为</span>局势的评分。<span style="color: black;">通常</span>到80%这类的胜率就没<span style="color: black;">道理</span>了,必胜了,<span style="color: black;">设备</span>看自己<span style="color: black;">小于</span>20%就中盘认输了。</span><span style="color: black;">AlphaGo的创新是有价值网络,<span style="color: black;">评定</span>叶子节点时不是只看下完的结果,而是一半一半,<span style="color: black;">亦</span><span style="color: black;">思虑</span>价值网络直接对叶子节点预测的胜负结果。走子<span style="color: black;">选取</span>就简单了,选获胜概率最大的那个分支。<span style="color: black;">设备</span><span style="color: black;">亦</span>会随机下,<span style="color: black;">由于</span>有时几个分支胜率<span style="color: black;">同样</span>。 )</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">人类<span style="color: black;">一般</span><span style="color: black;">运用</span>的是判断阵势,而AI能确定性的在大局上去“扣”细节。</strong></span>MCTS这个框架对棋力最大的<span style="color: black;">道理</span>,<span style="color: black;">便是</span>“大局观”好。无论局部<span style="color: black;">怎样</span>激烈战斗,所有的模拟都永远下完,全盘算子的个数。<span style="color: black;"><strong style="color: blue;"><span style="color: black;">况且</span>这个大局观从原理上就超过了人类!<span style="color: black;">例如</span>人看到<span style="color: black;">一起</span>阵势,<span style="color: black;">倘若</span>不是基本封闭的实空,到底价值多少<span style="color: black;">评定</span>起来其实是非常粗的。<span style="color: black;">能手</span>点目时经常<span style="color: black;">这般</span>,先把能点的目算清楚,有<span style="color: black;">有些</span>小阵势如无忧角就给个经验目数,<span style="color: black;">而后</span>加上贴目算双方精确目数的差值,这类估算有<span style="color: black;">非常多</span>不精确的<span style="color: black;">原因</span>。</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">优良</span>二:AlphaGo比其它程序强,<span style="color: black;">乃至</span>比职业<span style="color: black;">能手</span>还强的,是近身搏杀时的小手段。</span></strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><img src="http://mmbiz.qpic.cn/mmbiz/LaSSr9wZrk5T3Aprhxib6E1usJnK87icnclgGTp19Mw66a5mGDRfPeRZZdv5G06lB1VHbncndhF92CpdZicPkDpJg/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></strong><span style="color: black;">第三局,李世石29和31是失着。29凑白30双,虽然<span style="color: black;">得到</span>了H17的先手,<span style="color: black;">然则</span>中间的头更为重要。当黑31手飞出后,白32象步飞<span style="color: black;">能够</span>说直接将黑击毙了。在盘面的左上中间焦点处,AlphaGo的快速走子网络会有一个7*7之类的小窗口,对<span style="color: black;">这儿</span>进行穷举<span style="color: black;">同样</span>的搜索,用人手写的代码加上策略网络。32这步妙招可能<span style="color: black;">便是</span><span style="color: black;">这般</span>找出来的,李世石肯定<span style="color: black;">无</span>算到。<span style="color: black;">然则</span>AlphaGo是不怕麻烦的,就<span style="color: black;">始终</span>对着<span style="color: black;">这儿</span>算,比人<span style="color: black;">更易</span>看到黑三子的可怜<span style="color: black;">结果</span>。这个计算对人有些<span style="color: black;">繁杂</span>,<span style="color: black;">仅有</span>实力很强的<span style="color: black;">才可</span>想到算清楚,对AlphaGo<span style="color: black;">便是</span>小菜。李世石一招<span style="color: black;">不小心</span>就被技术性击倒了。AlphaGo对这种封闭局部的计算,是它超过人类的强项。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"> 汪汪的可能缺陷:开放式接触与劫争还需完善。</span></strong><span style="color: black;">咱们</span>在系列(一)中<span style="color: black;">说到</span>过,在开放式接触中<span style="color: black;">因为</span>可能性太多,阿尔法汪会采取“剪枝”的<span style="color: black;">办法</span>,<span style="color: black;">倘若</span>战斗会搞到很远去,它<span style="color: black;">亦</span>可能手数太多算不清,露出破绽。而<span style="color: black;">倘若</span>是在开局<span style="color: black;">或</span>中局封闭式局部有了劫争,<span style="color: black;">因为</span>要找劫,等于强制变<span style="color: black;">成为了</span>杀到全盘的开放度最大的开放式局面了。这是AlphaGo不<span style="color: black;">爱好</span>的,它的小窗口搜索就用不上了。而用MCTS搜索,打劫步数<span style="color: black;">太多</span>,就会超过它的叶子节点扩展深度,<span style="color: black;">例如</span>20步就不行了,<span style="color: black;">必要</span>“快速走子”收完了。<span style="color: black;">此时</span>它就胡乱终局了,不<span style="color: black;">晓得</span><span style="color: black;">怎样</span>处理劫争,模拟质量<span style="color: black;">快速</span>下降。<span style="color: black;"><strong style="color: blue;">从第四盘来观察,在“神78手”之后,整体局面<span style="color: black;">处在</span>不符合常规且多处争夺。此刻推测阿尔法汪的“剪枝效应”使得对局面的<span style="color: black;">评定</span><span style="color: black;">显现</span>错误,反而还不如人类的“大局观”,这一点在Demis Hassabis的回复中<span style="color: black;">亦</span>有验证,<span style="color: black;">因此</span>以核心价值观将各算法完美糅合的方式或许还需完善。</strong></span></p>
<span style="color: black;"><strong style="color: blue;"><span style="color: black;">写在后面,以及投资</span></strong></span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;"> 在系列(二)中<span style="color: black;">咱们</span>曾表达过对<span style="color: black;">全部</span>AI发展水平超乎<span style="color: black;">掌控</span>的担心,</span><span style="color: black;">从<span style="color: black;">Demis Hassabis的反馈来看虽然阿尔法汪<span style="color: black;">已然</span>很优秀,<span style="color: black;">咱们</span>仍然<span style="color: black;">能够</span>从算法本身的特征来分析阿尔法汪的<span style="color: black;">优良</span>与缺陷</span><span style="color: black;">(类似系列(一),<span style="color: black;">幸运</span>自己不是在完全瞎扯)。</span></span></span></strong></span></strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;"><span style="color: black;">AlphaGo算法里有<span style="color: black;">有些</span>模块代码是<span style="color: black;">必须</span>人去写的,这些代码可不是<span style="color: black;">设备</span>训练出来的。例如蒙特卡洛搜索(MCTS)<span style="color: black;">全部</span>框架的代码,例如快速走子网络的代码。其实有两位论文<span style="color: black;">一起</span><span style="color: black;">第1</span>作者David Silver和Aja Huang<span style="color: black;">数年</span><span style="color: black;">累积</span>的贡献。这些人写的代码,就会有内在的缺陷,是代码的缺陷,<span style="color: black;">乃至</span>可能是多种算法合成思想的缺陷,<span style="color: black;">咱们</span><span style="color: black;">亦</span>不必太过于神话。</span></span></span></strong></span></strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">针对</span>投资而言,<span style="color: black;">咱们</span><span style="color: black;">全部</span>系列的观点<span style="color: black;">始终</span>从一而终:<span style="color: black;"><strong style="color: blue;">AI是<span style="color: black;">将来</span>,但A股技术还太远,<span style="color: black;">商场</span>模式更未明。</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">(附系列(一)投资评论:</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"> <span style="color: black;">倘若</span>纯从AI技术能力来看,A股企业与Deep Mind尚有<span style="color: black;">必定</span>差距。科大讯飞<span style="color: black;">关联</span>实验室在国内属于领先水平,接近百度、港中文、港科大等实验室。<strong style="color: blue;">国内AI概念股<span style="color: black;">重点</span>停留在算法类似的语音识别、模式识别、<span style="color: black;">设备</span>视觉等技术,<span style="color: black;">商场</span>模式短期内<span style="color: black;">没法</span>大规模实现</strong>,<span style="color: black;">因此</span>各类概念股在<span style="color: black;">这儿</span>就不一一列举。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"> 从<span style="color: black;">这次</span>对抗来看,人工智能的两方面不足值得关注。一方面是前期计算需求大<span style="color: black;">引起</span>不得不舍弃许多模拟,另一方面是仍然<span style="color: black;">没法</span>完全复制人类的思考模型。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">计算的需求<span style="color: black;">能够</span>用云的方式<span style="color: black;">处理</span>,而<span style="color: black;">更加多</span>更<span style="color: black;">繁杂</span>的思考模型则是大数据发展的重要步伐,<span style="color: black;">亦</span>是人类的自我探索。在IT的发展路途中,人或许是对自我的模仿,云计算就像<span style="color: black;">咱们</span>把思考<span style="color: black;">拜托</span>在广泛的神经元,<span style="color: black;">而后</span>大数据<span style="color: black;">逐步</span>模拟人脑整合海量信息的方式,更美妙的是IT比人脑更准确更少出错。)</span></strong></p>
<span style="color: black;"><strong style="color: blue;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">联系人:刘洋、刘高畅 13920105213/刘智、刘畅、孙家旭等</span></span></strong></span></strong></span>
百度seo优化论坛 http://www.fok120.com/ 感谢您的精彩评论,为我带来了新的思考角度。 你的见解真是独到,让我受益匪浅。 真情实感,其含义为认真了、走心了的意思,是如今的饭圈常用语。
页:
[1]