AlphaZero 用自我博弈统一棋类学习

三张棋盘。三套从随机参数起步的权重。一套算法。

2017 年 12 月 5 日前后公开、后在 Science 完整发表的 AlphaZero，把问题收得很干净：给定游戏规则、自我对弈与足够计算，同一套方法能否分别精通国际象棋、将棋与围棋？答案必须分三次写——不是一个权重文件通吃三棋，而是三次独立训练。每一种游戏都不输入人类对局作监督标签；网络同时输出走子概率与局面价值；蒙特卡洛树搜索用当前网络作先验与评估，搜索得到的改进策略再当训练目标。棋盘表示、合法动作与胜负判定由规则给出。没有开局库，没有国际象棋引擎传统里那类手工评价函数，也没有为某一棋种特制的启发式长列表。

论文报告：在规定训练时间内，AlphaZero 在国际象棋上击败当时的 Stockfish 对阵设置，在将棋上击败 Elmo，在围棋上击败 AlphaGo Zero 的对阵版本。时限、硬件与对手版本以论文表格为准——受控实验，不是开放人类比赛，也不是把三张棋盘装进同一个神经网络。

AlphaGo 已经证明深度强化学习可以下好围棋，却仍带着人类棋谱的监督预训练，以及为围棋剪裁的管道。AlphaZero 拿掉的，是专家棋谱里的走法先验。“无人类知识”这句宣传语要拆开读：规则、对称性处理、搜索算法与训练超参仍是人类设计；被降级的是人类对局频率对走法好坏的主导权。对国际象棋社区，这一点尤其刺眼——引擎进步长期依赖评价函数与开局库的人工积累，而自我对弈生成的数据改写了风格评论者常提到的弃子与长期压迫频率。那些观察来自对局与评论，不宜夸大成神秘棋风本质；严谨比较还必须控制对局时间与硬件。

方法统一的意义大于任何一盘名局。在规则明确、可大量模拟、终局可判定的环境里，专家数据可以从门槛变成加速器；策略—价值网络与树搜索的闭环，把计算转化成对局经验，再把经验转化成更强的网络。机器人、对话或科学发现很少具备如此干净的模拟器。AlphaZero 不是通用智能的证明，而是一张关于“可模拟决策问题上限”的清晰样张：叙述可迁移，权重不共享；干净来自问题本身的干净。

随机初始化、自我对弈、策略价值一体网络、蒙特卡洛树搜索——清单不长，约束很强。强在规则与模拟器，不在口号里的“零知识”。读懂约束，才读得懂迁移边界。

Three boards. Three weight sets started from random parameters. One algorithm.

AlphaZero, public around 5 December 2017 and later published fully in Science, cleans the question: given rules, self-play, and enough compute, can one method separately master chess, shogi, and Go? The answer must be written three times—not one weight file spanning three games, but three independent training runs. No human games as supervised labels. A network outputs move probabilities and a position value. Monte Carlo tree search uses the network as prior and evaluator; improved policies from search become training targets. Board representation, legal actions, and terminal outcomes come from the rules. There is no opening book, no handcrafted evaluation function of the kind traditional chess engines rely on, and no long list of game-specific heuristics.

The paper reports that within set training budgets AlphaZero beat contemporaneous match setups against Stockfish in chess, Elmo in shogi, and a comparable AlphaGo Zero setup in Go. Time controls, hardware, and opponent versions are those of the paper’s tables—controlled experiments, not open human matches, and not three games stuffed into one neural network.

AlphaGo had already shown deep reinforcement learning could play Go at a high level, yet still carried supervised pretraining on human games and a pipeline tailored to one board. What AlphaZero removes is move priors from expert records. The slogan “no human knowledge” needs unpacking: rules, symmetry handling, search, and hyperparameters remain human designs; what is demoted is the dominance of human game frequencies over which moves count as good. For the chess community that cut especially deep—engine progress had long leaned on hand-built evaluation and books—while self-play data rewrote the frequencies of sacrifices and long-term pressure that commentators often notice. Those descriptions come from games and commentary; they should not be inflated into a mystical essence of style, and rigorous comparison still requires matched time and hardware.

Unity of method matters more than any single famous game. In environments with clear rules, cheap simulation, and decidable outcomes, expert data can fall from gate to accelerator. A closed loop of policy-value network and tree search turns compute into games and games into a stronger network. Robotics, dialogue, and scientific discovery rarely offer such clean simulators. AlphaZero is not a proof of general intelligence; it is a sharp specimen of what highly simulable decision problems allow—transferable narrative, unshared weights; cleanliness that belongs to the problem itself.

Random initialization, self-play, a joint policy-value network, Monte Carlo tree search: a short list under strong constraints. The strength is in rules and simulators, not in the slogan “zero knowledge.” Read the constraints, and the transfer boundary comes into focus.