游戏ai人工智能
国际象棋和象征性AI (Chess and Symbolic AI)
Today, AI based on deep learning and neural networks is taking the world by storm. However, many of the algorithms that guide our web searches and driving directions today are much older, rooted in what people now call “Good, Old Fashioned AI,” also known as “symbolic” AI, which was the primary form of AI from the 1950s through the 1990s. The eclipse of symbolic AI by deep learning is illustrated by two major milestones in AI history, each featuring the world’s top human player being beaten in a game by an AI system.
基于深度学习和神经网络公司的T ODAY，AI正在席卷全球。 但是，当今指导我们的网络搜索和行车路线的许多算法都已经过时了，植根于人们现在所说的“老式老式AI”(也称为“符号” AI)，这是AI的主要形式。 1950年代到1990年代。 人工智能历史上两个主要的里程碑说明了深度学习使象征性AI黯然失色的情况，每个里程碑都体现了世界顶级人类玩家在游戏中被AI系统击败的经历。
The 1997 defeat of world champion chess grandmaster Garry Kasparov by IBM’s Deep Blue computer was hailed as a triumphant watershed in technological history, akin to the moon landing, because it appeared to show that computers could beat humans at something once thought to be exclusive to us: thinking. The symbolic AI techniques DeepBlue used are now considered passé, especially for more complicated games such as Go, believed to have been invented in China 2,500 years ago. But in 2016, world Go champion Lee Sedol was defeated by Google DeepMind’s AlphaGo AI system. This event has been called China’s “Sputnik moment” by AI researcher and venture capitalist Kai-Fu Lee, who argues that this specific event led China to pour billions of dollars of money into AI research to catch up with, and potentially surpass, the United States. AlphaGo’s victory illustrates the rise of the new paradigm in AI, deep learning and neural networks, at the heart of the current AI revolution.
1997年，IBM的Deep Blue电脑击败了世界冠军国际象棋特级大师Garry Kasparov，被誉为技术史上的胜利分水岭，类似于登月，因为它似乎表明，在曾经被认为是我们独有的东西上，计算机可以击败人类。 ：思考。 现在，DeepBlue使用的象征性AI技术被认为是过时的，特别是对于Go等更复杂的游戏而言，据信这种游戏是在2500年前在中国发明的。 但是在2016年，世界围棋冠军Lee Sedol被Google DeepMind的AlphaGo AI系统击败。 AI研究人员和风险资本家李开复(Kai-Fu Lee)将该事件称为中国的“人造卫星时刻”。他认为，这一特定事件使中国向AI研究投入了数十亿美元，以赶上并有可能超越美国。状态。 AlphaGo的胜利说明了AI，深度学习和神经网络新范式的兴起，这是当前AI革命的核心。
Why have games such as chess and Go been so important in the history of AI? The pioneers of artificial intelligence, which included Herbert Simon, Alan Newell, John McCarthy and Marvin Minsky, viewed human intelligence through a tradition of Western philosophy going back to Aristotle. This masculine-gendered, Eurocentric view of intelligence-rooted in the Cartesian separation of mind from body-privileges the cerebral skills of logic, math, and problem-solving ability over bodily, emotional, social, and cultural forms of intelligence. If Reason (i.e. Logic) separated Man from Beast, then Logic must be the basis for intelligence, they thought.
为什么象棋和围棋这样的游戏在AI历史上如此重要？ 人工智能的先驱包括赫伯特·西蒙，艾伦·纽厄尔，约翰·麦卡锡和Maven·明斯基，他们通过追溯到亚里士多德的西方哲学传统来观察人类的智力。 这种以男性为中心的，以欧洲为中心的智力观，植根于笛卡尔的思维与身体的分离，使他们的智力，逻辑，数学和解决问题的能力超越了身体，情感，社交和文化形式的智力。 他们认为，如果理性(即逻辑)将人与野兽分开，那么逻辑必定是智慧的基础。
Many Western philosophers and mathematicians, from Blaise Pascal to George Boole and Bertrand Russell aspired either to make calculation or logic, which they equated with thought itself, more mathematically rigorous (more “formal”), or to take the next step, to mechanize it. Pascal himself built a calculating machine for this purpose, and this Western impulse culminated in the invention of the digital computer in the 20th century. For the AI pioneers of the 1950s and 1960s, playing games were seen as just another way that people displayed intelligence by solving problems. If AI researchers could emulate how players did this, they could then automate the process. “Game theory,” a branch of mathematics with applications to economics and warfare both, was founded by mathematician and computer pioneer John von Neumann, and provided optimization strategies and algorithms more broadly useful to computer science. AI pioneer Herbert Simon applied such theories to both computer science and to economics (the field in which he won the Nobel Prize). Thus, the idea that games could seriously model aspects of the real world was central to early computer science. In particular, since early computers had difficulty modeling the complexities of the real world, games provided a simpler “micro-world,” with boundaries and rules easily understood by computers, that made rapid progress possible in the 1960s.
从布莱斯·帕斯卡(Blaise Pascal)到乔治·布尔(George Boole)和伯特兰·罗素(Bertrand Russell)的许多西方哲学家和数学家都渴望进行计算或逻辑化，他们将其等同于思想本身，在数学上更加严格(更“形式化”)，或者下一步是对其进行机械化。 。 帕斯卡(Pascal)亲自为此目的建造了一台计算机，这种西方的冲动最终导致了20世纪数字计算机的发明。 对于1950年代和1960年代的AI先锋来说，玩游戏只是人们通过解决问题来展示智慧的另一种方式。 如果AI研究人员可以模仿玩家的操作方式，那么他们就可以自动化该过程。 数学家和计算机先驱约翰·冯·诺伊曼(John von Neumann)共同创立了“博弈论”，这是数学的一个分支，在经济学和战争领域均有应用。它提供了对计算机科学更广泛有用的优化策略和算法。 人工智能的先驱赫伯特·西蒙(Herbert Simon)将此类理论应用于计算机科学和经济学(他获得诺贝尔奖的领域)。 因此，游戏可以认真模拟现实世界的各个方面的想法在早期计算机科学中至关重要。 特别是，由于早期的计算机难以模拟现实世界的复杂性，因此游戏提供了一个更简单的“微观世界”，其边界和规则易于计算机理解，从而在1960年代取得了飞速发展。
Chess, in particular, has historically been seen in the West as the pinnacle of intellectual activity. It was an intellectual’s game, associated with logic and strategy. Think of Star Trek’s Mr. Spock, and you may picture him at his 3D chess board, outplaying his human opponents. Even in the 18th century, European elites were fascinated by the idea of machines that might play chess. Wolfgang von Kempelen became famous for his “Mechanical Turk,” an automatic chess machine built for the Austrian Empress Maria Theresa, which defeated Benjamin Franklin and Napoleon. Eventually, the Turk turned out to be fake, with a real man inside, but it nevertheless captured the imaginations of Edgar Allen Poe as well as Charles Babbage. This interest in chess as a marker of intelligence carried on in the mathematicians that helped define the theory of computing in the 20th century: Alan Turing, Claude Shannon, John von Neumann, Norbert Wiener, and of course the AI pioneers Herbert Simon, Alan Newell, and John McCarthy. For Newell and Simon, in particular, chess was an exemplary problem for AI particularly fitted to their preferred solution: search.
特别是国际象棋，在历史上一直被视为西方知识活动的顶峰。 这是一个知识分子的游戏，与逻辑和策略相关。 想想星际迷航的 Spock先生，您可能会在他的3D棋盘上想象他，胜过他的人类对手。 即使在18世纪，欧洲精英也对可能下棋的机器的想法着迷。 沃尔夫冈·冯·肯佩伦(Wolfgang von Kempelen)以其“机械土耳其人”而闻名，这是为奥地利女皇玛丽亚·特蕾莎(Maria Theresa)建造的自动下棋机，击败了本杰明·富兰克林和拿破仑。 最终，土耳其人被证明是伪造的，里面有一个真正的男人，但它仍然吸引了埃德加·艾伦·坡(Edgar Allen Poe)和查尔斯·巴贝奇(Charles Babbage)的想象。 对象棋的兴趣是在帮助定义20世纪计算理论的数学家中进行的一种智力标记：艾伦·图灵，克劳德·香农，约翰·冯·诺伊曼，诺伯特·维纳，当然还有AI的先驱赫伯特·西蒙，艾伦·纽厄尔以及约翰·麦卡锡(John McCarthy)。 特别是对于Newell和Simon，国际象棋是AI的一个典型问题，特别适合他们的首选解决方案：搜索。
Related: Visit the CHM online exhibit Mastering the Game: A History of Computer Chess
相关：访问CHM在线展览“ 掌握游戏：计算机象棋的历史”
寻找解决方案 (Searching for a Solution)
What is search, and how can it be used to play chess? By search in the AI context, I don’t mean searching for text on the web using Google (although a web search engine may use search in the AI sense.) Instead, in AI, search refers to the trial and error process of traversing possible paths to solving a problem. Search is one of the fundamental methods in classical AI, also known as “symbolic” AI because these methods involve the manipulation of lists of symbols, such as in the solving of algebra problems. All kinds of problem-solving tasks, such as proving theorems, solving puzzles, playing games, and navigating a maze, involve making choices about what to try first. These choices can be modeled as a tree of branching decisions.
什么是搜索，如何将其用于下棋？ 在AI上下文中进行搜索并不是指使用Google在网络上搜索文本(尽管Web搜索引擎可能会使用AI进行搜索。)相反，在AI中，搜索是指遍历的反复试验过程解决问题的可能途径。 搜索是经典AI(也称为“符号” AI)的基本方法之一，因为这些方法涉及符号列表的操纵，例如在代数问题的求解中。 各种解决问题的任务，例如证明定理，解决难题，玩游戏以及在迷宫中导航，都涉及选择首先尝试的方法。 这些选择可以建模为分支决策树。
极小值 (Minimax)
Branching decision tree of a tic-tac-toe game. Like chess and many other 2 player games, tic-tac-toe can be played with the minimax algorithm. 井字游戏的分支决策树。 像国际象棋和许多其他2人游戏一样，井字游戏可以使用minimax算法来玩。For instance, let’s say we wanted to build a robot mouse searching for the way out of a maze (something Claude Shannon did in 1950). If it arrives at a 4-way intersection, it can go right, forward, or left if it is not allowed to go backwards. This gives it 3 possible choices. Computer scientists would say that it has a “branching factor” of 3. A basic way to program a computer to solve the maze would be to try out each of the choices, or branches, in turn. This is known as “brute force” search: try every choice. However, our mouse would undoubtedly run into another intersection before it had a chance to backtrack to try the other choices in its first intersection. Each time it reaches another intersection, it can choose between another 3 paths. We could specify how many intersections deep the mouse should search before backtracking and trying out a different path.
例如，假设我们要构建一个机器人鼠标来寻找迷宫的出路(Claude Shannon在1950年所做的事情)。 如果到达4路交叉路口，则不允许向后行驶，可以向右，向前或向左行驶。 这给了它3种可能的选择。 计算机科学家会说它的“分支因子”为3。对计算机进行编程以解决迷宫的基本方法是依次尝试每个选择或分支。 这就是所谓的“强力搜索”：尝试所有选择。 但是，我们的鼠标无疑有机会碰到另一个路口，才有机会回溯尝试在其第一个路口尝试其他选择。 每次到达另一个路口时，都可以在其他3条路径之间进行选择。 我们可以指定在回溯并尝试其他路径之前，鼠标应搜索多少个交叉点。
This is known as the search depth or in the context of games, “look ahead.” As you can see, the number of paths that the mouse needs to search gets very big very quickly: 3, or the branching factor, multiplied by itself the number of times we look ahead in the decision tree. In other words, the problem grows exponentially. This is often called the “combinatorial explosion” problem in AI.
这就是所谓的搜索深度，或者在游戏中称为“向前看”。 如您所见，鼠标需要搜索的路径数量很快变得非常大：3，即分支因子乘以自身，我们在决策树中向前看的次数。 换句话说，问题成倍增长。 这通常被称为AI中的“组合爆炸”问题。
早期理论的棋盘 (A Chessboard of Early Theories)
Similar methods can be used for chess. Every player’s turn, they have a choice of up to 38 possible legal moves, giving the chess problem a branching factor of 38. A quantitative method of evaluating the relative advantage of one chess position over another is used to choose the best move out of these 38. This is called the “evaluation function.” A typical chess game takes about 42 moves, and because there are two players, this is multiplied by two, giving roughly 38⁸⁴ possible choices for the entire game, which is roughly 10³⁴, greater than the number of stars in the universe. Very early in the history of AI, it was recognized that brute force search of chess and other problems simply would not work with the hardware of the time; there were too many possibilities to consider, and computers were too slow. Claude Shannon, who was among the first to use the “ minimax “ algorithm in a computer chess program (an algorithm that is still the basis of most chess programs today), noted that one could apply human knowledge and experience of the game to quickly eliminate many branches from consideration. Herbert Simon and Alan Newell proposed the use of “heuristics,” or rules of thumb that humans often use in problem solving, that work most of the time but not always. Heuristics are a form of human knowledge that can be programmed into a computer.
类似的方法可以用于国际象棋。 每位玩家回合，他们最多可以选择38种可能的合法举动，从而使国际象棋问题的分支因子为38。一种评估一种国际象棋棋盘相对于另一国际象棋棋盘的相对优势的定量方法用于从这些棋子中选择最佳棋局。 38.这称为“评估功能”。 一个典型的国际象棋游戏需要大约42个动作，并且由于有两个玩家，所以乘以2，整个游戏中大约有38种可能的选择，这大约比宇宙中的星星数大10³。 在AI历史的早期，人们就认识到用强力搜索象棋和其他问题根本无法解决当时的硬件问题。 有太多的可能性要考虑，并且计算机速度太慢。 克劳德·香农(Claude Shannon)是最早在计算机国际象棋程序中使用“ maxmax ”算法的人(该算法仍是当今大多数国际象棋程序的基础)，他指出，人们可以运用人类知识和游戏经验来快速消除许多分支机构都在考虑之中。 赫伯特·西蒙(Herbert Simon)和艾伦·纽厄尔(Alan Newell)提出了“启发式”的使用方法，即人类在解决问题中经常使用的经验法则，这种方法在大多数情况下有效，但并非总是如此。 启发式是人类知识的一种形式，可以将其编程到计算机中。
棋 (Chess)
Chess has many more branches in its decision tree than tic-tac-toe. The number of chess games is estimated to be 10¹²⁰, more than the number of atoms in the universe. 国际象棋在决策树中的分支比井字游戏要多。 象棋游戏的数量估计为10十二⁰，比宇宙中原子的数量还要多。有界的前瞻 (Bounded Lookahead)
Given the enormous number of branches, chess programs can only look ahead to a finite depth in the search tree or be overwhelmed. 鉴于分支机构的数量众多，国际象棋程序只能在搜索树中前瞻有限的深度，否则将不知所措。One such heuristic that was found to be useful in chess was known as “ alpha-beta pruning.” This technique meant that if one of your moves was found to be easily countered by your opponent, other ways they could counter the same move didn’t have to be found. Further search along this path could be ignored, pruning off this entire branch of the decision tree. This could greatly reduce the branching factor, from 38 down to 6 or sometimes as low as 3. In addition, given the limits of computers at the time, most programs could only look 4 moves ahead. One of the earliest chess programs to be able to play competently against amateurs was created between 1959 and 1962 by MIT students including Alan Kotok, being advised by John McCarthy. The Kotok-McCarthy program used alpha-beta pruning.
一种被认为在国际象棋中有用的启发式方法被称为“ alpha-beta修剪” 。 该技术意味着，如果发现您的一个举动很容易被对手反击，则无需寻找其他可以反击相同举动的方式。 沿着该路径的进一步搜索可能会被忽略，从而删掉了决策树的整个分支。 这样可以大大降低分支因子，从38降低到6或有时低至3。此外，由于当时的计算机数量有限，大多数程序只能向前推进4个。 约翰·麦卡锡(John McCarthy)建议，麻省理工学院的学生，包括艾伦·科托克(Alan Kotok)，在1959年至1962年间创立了最早的能够与业余选手进行对抗的国际象棋程序之一。 Kotok-McCarthy程序使用了alpha-beta修剪。
艾伦·科托克(Alan Kotok)口述历史要点 (Highlights of Alan Kotok Oral History)
In 1959, MIT freshmen Alan Kotok, Elwyn R. Berlekamp, Michael Lieberman, Charles Niessen, and Robert A. Wagner started working on a chess-playing program based on research by artificial intelligence pioneer professor John McCarthy. By the time they had graduated in 1962, the program could beat amateurs. Collection of the Computer History Museum, 102645440. 1959年，麻省理工学院的新生Alan Kotok，Elwyn R. Berlekamp，Michael Lieberman，Charles Niessen和Robert A. Wagner开始着手一项基于人工智能先驱教授John McCarthy的研究的国际象棋计划。 到1962年他们毕业时，该程序已经可以击败业余爱好者了。 102645440，计算机历史博物馆收藏。Newell and Simon believed that all AI problems, like chess, could be solved by search combined with heuristics, or “heuristic search.” Heuristic search was the central idea behind Newell and Simon’s early breakthroughs, the Logic Theorist and the General Problem Solver, and was a key pillar in their theory that intelligence in both humans and machines was simply the manipulation of symbols, the fundamental building blocks of both mathematics and language. This “physical symbol system” hypothesis was the assumption upon which the entire project of symbolic AI was founded, from its beginnings in the 1950s through the early 2000s. This theory, which posited that computers and human minds were thus equivalent, became extremely influential in cognitive psychology as well, and ultimately made its way into popular culture through cyberpunk science fiction, in which humans can upload their brains to the internet or have them replaced by chips.
Newell和Simon认为，可以通过结合搜索法或“启发式搜索”来解决象棋这样的所有AI问题。 启发式搜索是纽维尔和西蒙早期突破背后的核心思想，即逻辑理论家和一般问题解决者，并且是他们理论中的关键Struts，即人类和机器中的智能仅仅是符号的操纵，是符号和符号的基本构建块数学和语言。 这个“物理符号系统”假说是从1950年代开始到2000年代初期整个符号AI项目建立的假设。 该理论假定计算机和人类的思想是等价的，它也对认知心理学产生了极大的影响，并最终通过赛博朋克科幻小说进入了大众文化，人们可以将他们的大脑上传到互联网上或替换它们。通过筹码。
Related: Explore the Turing Award Lecture by Allen Newell and Herbert Simon in 1975 that laid out their Physical Symbol System hypothesis and heuristic search as the fundamental basis for all symbolic AI
相关：探索1975年艾伦·纽厄尔和赫伯特·西蒙的图灵奖演讲 ，阐述了他们的物理符号系统假设和启发式搜索作为所有符号AI的基础
As computers got faster and as computer scientists who were also experienced chess players, such as Richard Greenblatt and Hans Berliner, created their own chess programs, they found that earlier chess programs (such as Kotok’s) played exceedingly poorly, and added their own knowledge of how human players approach the game to their programs in the form of additional heuristics to better evaluate chess positions, databases of opening moves and endgames, and recognizers of board patterns. However, over time, it turned out that chess programs running on faster computers or with custom hardware could beat programs with a lot of human knowledge built in. This was because no heuristic is perfect or can cover every case. Occasionally genius moves can occur if a player tries something that looks to most people like a bad move. Most heuristics would prune out such a move before searching it, and thus the canned-knowledge-based programs would never find such a move.
随着计算机变得越来越快，同时也有资深棋手(例如Richard Greenblatt和Hans Berliner)的计算机科学家创建了自己的棋子程序，他们发现较早的棋子程序(例如Kotok's)的玩法极其糟糕，并增加了自己的知识。人类玩家如何以其他试探法的形式将游戏推向他们的程序，以更好地评估棋位置，开局动作和残局的数据库以及棋盘图案的识别器。 但是，随着时间的流逝，事实证明，运行在速度更快的计算机上或具有自定义硬件的国际象棋程序可能会击败内置了很多人类知识的程序。这是因为没有一种试探法能完美地解决所有问题。 如果玩家尝试某些对大多数人来说都像坏举动的事情，有时就会发生天才举动。 大多数启发式算法会在搜索之前先修剪掉这样的动作，因此基于罐头知识的程序永远不会找到这样的动作。
As computers got faster, they could begin to look ahead more deeply, to 6, 7, 8 moves ahead, easily trouncing programs that only looked 4 moves ahead. A more efficient search algorithm, known as “iterative deepening search,” was discovered which could gradually increase the search depth of an avenue that looked promising. This was first used in David Slate and Larry Atkins’ Chess 4.5, which was the first program to win in a human chess tournament in 1976. More memory also allowed programs to save previously considered positions, further reducing the amount of search required. All of these innovations (alpha-beta pruning, iterative deepening, saving of searched positions, and databases of opening and endgames) were freely shared among chess program developers at computer chess tournaments, and became standard tricks of the trade.
随着计算机变得越来越快，他们可以开始更深入地向前看，向前前进6、7、8，很容易将仅向前看4的程序删节。 发现了一种更有效的搜索算法，称为“迭代加深搜索”，可以逐渐增加看起来很有希望的途径的搜索深度。 这是在David Slate和Larry Atkins的国际象棋4.5中首次使用的，该程序是1976年在人类国际象棋比赛中获胜的第一个程序。更多的内存也使程序可以保存以前考虑的位置，从而进一步减少了搜索量。 所有这些创新(alpha-beta修剪，迭代加深，保存搜索到的位置以及开局和残局数据库)在计算机国际象棋比赛中由国际象棋程序开发人员自由共享，并成为行业的标准把戏。
Despite these advances in software, as computers got faster in the 1970s, chess programs automatically got better without any software innovation. By the 1980s progress in computer chess was dominated by using hardware to accelerate search. It had become a computer design problem, not an AI problem. In 1997, Deep Blue was still using mostly the same software techniques as chess programs 20 years earlier, but beat Kasparov mainly by being a faster computer with many custom parallel processors. In a way, as computers got faster, the chess programs became less intelligent.
尽管软件取得了这些进步，但随着1970年代计算机的发展，象棋程序会自动变得更好，而无需任何软件创新。 到1980年代，计算机国际象棋的进步已被使用硬件加速搜索所主导。 它已经成为计算机设计问题，而不是AI问题。 1997年，Deep Blue仍使用与20年前象棋程序相同的软件技术，但击败Kasparov主要是因为它是一款具有许多自定义并行处理器的速度更快的计算机。 在某种程度上，随着计算机变得越来越快，国际象棋程序变得越来越不智能。
Deep search as a dominant theme in AI was already in decline in the 1980s. Beginning in the 1960s, researchers such as Ed Feigenbaum of Stanford were creating so-called “expert systems,” in which a lot of expert human knowledge was put into AI programs in the form of if-then rules. Like earlier heuristic programs, these rules were pre-programmed into the software, but unlike those systems, they were separated out into “knowledge bases” from the logical parts of the program, the “inference engines.” Feigenbaum and other expert systems proponents insisted that “in the knowledge lies the power.” In other words, a large knowledge base would make up for the lack of sophisticated reasoning: more knowledge meant less search, and vice versa.
深度搜索作为AI的主要主题，在1980年代已经开始下降。 从1960年代开始，斯坦福大学的Ed Feigenbaum等研究人员正在创建所谓的“专家系统”，其中，许多专家知识以if-then规则的形式输入到AI程序中。 像早期的启发式程序一样，这些规则已预先编程到软件中，但是与那些系统不同，它们被从程序的逻辑部分(即“推理引擎”)分离为“知识基础”。 费根鲍姆(Feigenbaum)和其他专家系统的支持者坚持认为，“知识在于力量”。 换句话说，庞大的知识库将弥补缺乏复杂推理的不足：更多的知识意味着更少的搜索，反之亦然。
老虎在笼子里：基于知识的系统的应用，Edward Feigenbaum的演讲，1993年 (Tiger in a Cage: Applications of Knowledge-Based Systems, lecture by Edward Feigenbaum, 1993)
AAAI-17 AI历史受邀小组：专家系统，2017年 (AAAI-17 Invited Panel on AI History: Expert Systems, 2017)
Related: Read Expert Systems Industry Workshop Transcripts, 2018
相关：阅读专家系统行业研讨会成绩单，2018年
Expert systems spawned many commercial companies in the 1980s. Little of this activity affected chess programs, which at this time were turning the opposite direction: back to brute force search through the use of dedicated hardware. The leading chess machines of this type were Ken Thompson’s Belle from Bell Labs, and two separate projects from Carnegie Mellon, Hans Berliner’s HiTech and Feng-Hsiung Hsu and Murray Campbell’s Deep Thought, which later became IBM’s Deep Blue. By the time of Kasparov’s defeat, then, chess programs had already become largely irrelevant to the greater field of AI, even if they provided some good PR.
专家系统催生了1980年代的许多商业公司。 这项活动几乎没有影响国际象棋程序，而国际象棋程序此时正朝着相反的方向发展：通过使用专用硬件回到暴力搜索。 这种类型的领先国际象棋机器是Bell Labs的Ken Thompson的Belle，以及Carnegie Mellon，Hans Berliner的HiTech和Feng-Hsiung Hsu的两个独立项目以及Murray Campbell的Deep Thought，后来变成了IBM的Deep Blue。 因此，到卡斯帕罗夫(Kasparov)被击败时，国际象棋程序已经与更大的AI领域无关，即使它们提供了良好的PR。
More worrisome, however, was that by the 1990s, the symbolic AI project based on Newell and Simon’s physical symbol system hypothesis was under attack. Critics, such as the philosopher Hubert Dreyfus, had questioned the symbolic AI project as early as the 1960s, on the basis that the philosophical assumption of mind-body separation that symbolic AI rested on was incorrect and out of date. 20th century philosophers such as Martin Heidegger had insisted that human thought could not be separated from the experience of the body or from one’s immediate cultural surroundings.
然而，更令人担忧的是，到1990年代，基于Newell和Simon的物理符号系统假设的符号AI项目正受到攻击。 诸如哲学家休伯特·德雷福斯(Hubert Dreyfus)之类的批评者早在1960年代就对象征性AI计划提出了质疑，其依据是象征性AI所基于的身心分离的哲学假设是不正确的和过时的。 像马丁·海德格尔(Martin Heidegger)这样的20世纪哲学家坚持认为，人类的思想不能与身体的经验或与自己的直接文化环境相分离。
Related: Alchemy and Artificial Intelligence Report, Dreyfus, Hubert L., 1965
相关：《 炼金术与人工智能报告》，德雷福斯，休伯特·L，1965年
Related: What Computers Still Can’t Do, Dreyfus, Hubert L., 1992, 2nd ed. of What Computers Can’t Do, originally published 1972, revised 1979
相关内容： 《计算机仍然无法做什么》，德雷福斯，休伯特·L，1992年，第二版。 《计算机不能做什么》，最初于1972年出版，1979年修订
The reaction to Dreyfus’ critiques from AI researchers was savage (though Dreyfus was himself was not very diplomatic), with leading AI figures threatening journals if they published anything from him. They gloated when Dreyfus, who was not a particularly good chess player, was beaten by Richard Greenblatt’s chess program MacHack. Yet success at chess did not disprove Dreyfus’ critiques. In fact, it was the very fact that chess programs such as Deep Blue used brute force search that made it irrelevant in the larger project of creating general AI. The drama of Kasparov’s stunning defeat may have been hyped as a landmark in Machines triumphing over Man, but in reality it was a triumph of the Deep Blue engineers over one chess player. And Deep Blue’s creators did not claim the computer was intelligent. If the building was on fire, they said, Kasparov would be smart enough to run out but the machine would still sit there. And though in earlier years AI pioneer John McCarthy had thought chess a central problem of AI, after Deep Blue he criticized chess as not having developed any new theories for how human intelligence might be imitated.
AI研究人员对Dreyfus批评的React是野蛮的(尽管Dreyfus本人不是很外交)，如果AI领先人物发表任何论文，就会威胁到期刊。 当不是特别出色的棋手的德雷福斯被理查德·格林布拉特的棋子程序MacHack击败时，他们感到非常高兴。 然而，国际象棋的成功并不能反驳德雷福斯的批评。 实际上，正是像Deep Blue这样的国际象棋程序使用强力搜索的事实，使得它与创建通用AI的较大项目无关。 卡斯帕罗夫(Kasparov)惨败的戏剧可能被认为是《机器争霸》(Machines)战胜人类的里程碑，但实际上，这是《深蓝》(Deep Blue)工程师战胜一名国际象棋棋手的胜利。 Deep Blue的创建者并未声称计算机是智能的。 他们说，如果建筑物着火了，卡斯帕罗夫将足够聪明，可以用完，但机器仍会停在那里。 尽管在前些年，人工智能先驱约翰·麦卡锡(John McCarthy)认为国际象棋是人工智能的核心问题，但在《深蓝》(Deep Blue)之后，他批评国际象棋尚未开发出任何关于如何模仿人类智力的新理论。
By the 1990s, researchers were taking Dreyfus’ critiques seriously and imagining new forms of AI, such as those that emphasized having a body, like the robots of Rodney Brooks, or those that dealt with emotions. And as we shall see in Part 2, in the 2000s a completely different tradition of AI, called machine learning, would start to replace symbolic AI as the way forward. Machine learning would perform feats that symbolic AI had never been able to tackle better than humans, such as recognize faces or understand human speech. This would also apply to a game that had been infeasible to play competitively using heuristic search: Go.
到1990年代，研究人员开始认真对待Dreyfus的批评，并想象新的AI形式，例如那些强调拥有身体的机器人，例如Rodney Brooks的机器人，或者那些处理情绪的机器人。 正如我们将在第2部分中看到的那样，在2000年代，一种完全不同的AI传统(称为机器学习)将开始取代符号AI。 机器学习将实现象征性AI从未能够比人类更好地应对的壮举，例如识别面部表情或理解人类语音。 这也适用于使用启发式搜索无法竞争的游戏：Go。
Nevertheless, despite search losing its luster as the central technique of AI, it has never lost its usefulness in computer science more broadly. There has been much progress in innovating better search algorithms over the years to solve problems optimally and efficiently. Being such a fundamental technique, the creation of decision trees and searching through them is so widespread that it is likely impossible to get an accurate accounting of all the programs that use it today.
尽管如此，尽管搜索失去了作为AI核心技术的光彩，但它从未在更广泛的计算机科学领域失去它的用处。 多年来，在创新更好的搜索算法以优化和有效解决问题方面取得了很大进展。 作为一种基本技术，决策树的创建和搜索通过它们是如此广泛，以至于不可能对当今使用它的所有程序进行准确的计算。
Search plays some role in any information retrieval task, from querying a database to searching the Web. The A* search algorithm, which was first invented for SRI’s robot Shakey, is commonly used for route-finding in autonomous vehicles and GPS apps. And even today, game playing AI programs that use machine learning utilize some form of search, even if it is no longer the most interesting component. However, like other techniques that used to be considered “AI,” today search may be seen merely as just another basic computer technique, no more intelligent than the next program. This illustrates a historical pattern in the development of AI: once it becomes commonplace and automatic, humans no longer consider it “intelligence.” In years past, when one said “AI,” search was probably involved. When a person says “AI” today, however, they usually mean symbolic AI’s successor, machine learning.
从查询数据库到搜索Web，搜索在任何信息检索任务中都扮演着重要角色。 A *搜索算法最早是为SRI的机器人Shakey发明的，通常用于自动驾驶汽车和GPS应用中的路线查找。 即使在今天，使用机器学习的玩游戏AI程序也会使用某种形式的搜索，即使它不再是最有趣的组成部分。 但是，就像以前被认为是“ AI”的其他技术一样，今天的搜索可能仅被视为另一种基本计算机技术，没有比下一个程序更智能。 这说明了AI发展的历史模式：一旦AI变得司空见惯和自动化，人们就不再将其视为“智能”。 在过去的几年中，当人们说“ AI”时，可能涉及搜索。 但是，当今天一个人说“ AI”时，它们通常是象征性AI的继任者，即机器学习。
了解AI的深度学习革命，深度学习与搜索和符号AI的不同之处，以及DeepMind的AlphaGo如何利用深度学习在AI和Play的第2部分：Go和深度学习中击败世界冠军围棋选手Lee Sedol。 (Learn about the deep learning revolution in AI, how deep learning differs from search and symbolic AI, and how DeepMind’s AlphaGo utilized deep learning to beat Lee Sedol, the world champion Go player, in AI and Play, Part 2: Go and Deep Learning.)
笔记 (Notes)
1. Nathan Ensmenger, “Is Chess the Drosophila of Artificial Intelligence? A Social History of an Algorithm,” Social Studies of Science 42, no. 1 (February 2012): 22, https://doi.org/10.1177/0306312711424596.
1. Nathan Ensmenger，“国际象棋是果蝇吗？” 算法的社会史》，《 科学社会研究》第 42期，第1期。 1(2012年2月)：22， https ： //doi.org/10.1177/0306312711424596。
2. Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order. (Boston; New York: Houghton Mifflin Harcourt, 2019), 1–5.
2.李开复， 人工智能超级大国：中国，硅谷和新世界秩序。 (波士顿；纽约：霍顿·米夫林·哈科特，2019年)，第1至5页。
3 Stuart J. Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 3rd ed., Prentice Hall Series in Artificial Intelligence (Upper Saddle River, NJ: Pentice Hall, 2010), 110.
3 Stuart J.Russell和Peter Norvig，《 人工智能：一种现代方法》 ，第三版，《人工智能的Prentice Hall系列》(新泽西州上萨德尔河，Pentice Hall，2010年)，第110页。
4 Rodney A. Brooks, “Elephants Don’t Play Chess,” Robotics and Autonomous Systems, Designing Autonomous Agents, 6, no. 1 (June 1, 1990): 3–15, https://doi.org/10.1016/S0921-8890(05)80025-9.
4 Rodney A. Brooks，“大象不要下象棋”， 机器人与自治系统 ，设计自治代理，第6号，第1期。 1(1990年6月1日)：3-15， https ：//doi.org/10.1016/S0921-8890(05)80025-9 。
从CHM探索更多 (Explore More From CHM)
展品 (Exhibits)
Mastering the Game: A History of Computer Chess
掌握游戏：计算机国际象棋的历史
Artificial Intelligence gallery, Revolution: The First 2000 Years of Computing
人工智能画廊，“革命：2000年的第一年计算”
大事记 (Events)
CHM Live│The History of Computer Chess: An AI Perspective — Panel discussion with Murray Campbell, Ed Feigenbaum, David Levy, and John McCarthy. Moderated by Monty Newborn, 2005
CHMLive│计算机国际象棋的历史：人工智能的视角—与Murray Campbell，Ed Feigenbaum，David Levy和John McCarthy进行的小组讨论。 由Monty Newborn主持，2005年
口述历史 (Oral Histories)
Alan Kotok
艾伦·科托克(Alan Kotok)
Richard Greenblatt
理查德·格林布拉特
Monty Newborn
蒙蒂·新生儿
Harry Nelson
哈里·尼尔森
Ken Thompson
肯·汤普森
Peter Jennings
彼得·詹宁斯
Kathleen and Danny Spracklen
凯瑟琳和丹尼·斯普拉克伦
David Levy
大卫·利维
Hans Berliner
汉斯·柏林
Feng-Hsiung Hsu
徐凤雄
Murray Campbell
默里·坎贝尔
Originally published at https://computerhistory.org on July 23, 2020.
最初于 2020年7月23日 在 https://computerhistory.org 上 发布 。
翻译自: https://medium.com/chmcore/ai-and-play-part-1-how-games-have-driven-two-schools-of-ai-research-80e32534b043
游戏ai人工智能