AlphaGo 击败李世石

4 比 1。

2016 年 3 月，DeepMind 的 AlphaGo 在首尔与李世石完成五局赛，总比分落成这两位数字。赛前对照已经足够刺眼：AlphaGo 曾以 5 比 0 击败欧洲冠军樊麾；李世石则拿过十八个世界冠军头衔。围棋的合法局面数量极大，穷举走不通。系统必须先判断哪些落子值得算，再估计尚未终局的盘面大概对谁有利。

Nature 论文描述的架构把这两种判断分给不同网络。策略网络从职业棋谱的监督学习与自我对弈中学习落子分布，缩小候选；价值网络估计当前局面胜率；蒙特卡洛树搜索在二者引导下分配计算，把模拟集中到更有希望的变化。训练流程大致是：先模仿人类，再通过自我对弈强化，搜索产生的对局又反馈给网络。它既不是一本固定棋谱，也不是单次神经网络直觉落子。学习提供方向，搜索负责校验。

前三局 AlphaGo 连胜。第二局第 37 手（AlphaGo 执黑）偏离职业棋手的常规优先级，赛后被反复复盘：系统从人类数据起步，却在自我对弈与搜索中把低先验的落点推上了实战棋盘。程序仍以自己的评估函数工作；职业棋手愿意认真计算的范围，从此多了一个坐标。

第四局方向倒转。李世石执白第 78 手同样极不寻常，迫使 AlphaGo 进入罕见局面并连续失误，李世石拿下整场比赛唯一胜局。这局把系统边界写得很清楚：整体胜率优势并不抹去分布之外的脆弱；一次足够陌生的局部，仍可能让评估与搜索一起偏离。第五局由 AlphaGo 获胜，比分定格。

比赛使用中国规则，每方用时受限，AlphaGo 跑在分布式计算资源上，本地与云端机器协同完成搜索。公开材料描述了策略网络、价值网络的训练阶段，以及滚动出数百万局自我对弈的数据规模。第 37 手与第 78 手之所以被单独记住，是因为它们把“分布外落子”这一抽象概念，落成了可以在棋谱软件里逐步点击的坐标。

公众叙事喜欢把五局说成时代交替。棋手与工程师更常回看两个编号——37 与 78。一手来自系统，一手来自李世石，都落在既有经验不愿优先考虑的地方，也分别暴露了对手原先看不见的判断边界。围棋没有在穷举意义上被“解决”；被展示的是学习与搜索的分工，以及自我生成对局数据如何扩展人类棋谱覆盖不到的区域。后来的 AlphaZero 会进一步拿掉对人类棋谱的依赖；首尔这五局，仍停在人机共享棋盘、互相改写直觉的时刻。

四比一不是零封。第三十七手与第七十八手也不是宣传口号，而是棋谱坐标——可被点击、可被争辩、可被后续版本改写。公开五局把“策略网络缩小候选，价值网络估计胜率，树搜索消耗计算”焊进了公共记忆；焊点不在口号，而在两手可以定位的落子。

4–1.

In March 2016 DeepMind’s AlphaGo finished a five-game match against Lee Sedol in Seoul with that score. The pre-match contrast was already sharp: AlphaGo had beaten European champion Fan Hui 5–0; Lee held eighteen world titles. Go’s state space is too large for exhaustive search. A program must decide which moves deserve calculation and estimate who is ahead before the game ends.

The Nature architecture splits those judgments. A policy network, trained first on professional games and then with self-play, proposes move distributions and narrows candidates. A value network estimates winning probability from a position. Monte Carlo tree search allocates compute under both signals, concentrating simulations on promising lines. The pipeline roughly imitates humans, then strengthens through self-play, feeding search-generated games back into the networks. It is neither a fixed book nor a single neural intuition drop. Learning points; search checks.

AlphaGo took the first three games. Move 37 in game two—AlphaGo playing Black—departed from ordinary professional priorities and was replayed for years afterward. The system began from human data, yet self-play and search pushed a low-prior move onto a real board. The program still worked by its own evaluations; the set of lines professionals bother to read carefully gained a new coordinate.

Game four reversed the direction. Lee’s move 78 as White was equally unusual, forcing AlphaGo into a rare position and a cascade of errors. Lee took his only win of the match. The game writes the system boundary clearly: a global win-rate edge does not erase fragility outside the training distribution; a sufficiently strange local fight can still knock evaluation and search off together. AlphaGo won game five; the scoreboard froze.

Match conditions used Chinese rules and limited time; AlphaGo ran on distributed compute coordinating local and cloud machines. Public materials describe policy and value training stages and self-play at enormous scale. Move 37 and move 78 endure because they turn the abstract idea of out-of-distribution moves into coordinates one can click through in game software.

Public narrative likes to call the five games a changing of eras. Players and engineers more often return to two numbers—37 and 78. One move from the system, one from Lee, both landing where prior experience preferred not to look first, each exposing a judgment edge the opponent had not seen. Go was not “solved” by enumeration. What was shown was a division of labor between learning and search, and how self-generated games extend regions human game records never cover. AlphaZero would later drop the dependence on human games; Seoul still sits at the moment when human and machine shared one board and rewrote each other’s intuitions.

4–1 is not a shutout. Moves 37 and 78 are not slogans; they are coordinates—clickable, arguable, revisable by later versions. The public five games welded “policy proposes, value estimates, search spends compute” into memory at two places on a board.