关于「策略熵:定式集中度、局部最优与从众锁定」,《围棋 AI 后的人类决策质量》给出的定量结论是什么?

结论:第四章用香农熵度量定式集中度,示意角部局面的策略熵从 1.608 升到 2.360 比特,从众学习模型显示外部评估者让群体提早约 12 年离开局部最优。

数据口径:书中除论文数字外的参数均为示意,全部计算由书中程序生成;本书是非个人化经营建议,不针对任何个人决定。 第四章用香农熵度量定式集中度,示意角部局面的策略熵从 1.608 升到 2.360 比特,从众学习模型显示外部评估者让群体提早约 12 年离开局部最优。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 1.608, 2.360

era H bits MM bits2^H 95 % bootstrap CI top share
pre1.608 1.612 3.05 [1.517, 1.690] 0.620
post2.360 2.364 5.13 [2.313, 2.397] 0.300
maximum for 6 options2.585 bits
Jensen-Shannon divergence pre vs post0.0928 bits
small samples from the post-AI law (300 draws each):
n mean plug-in H mean Miller-Madow H
101.907 2.149
302.238 2.351
1002.323 2.359
10002.357 2.360
查看完整验算程序
# Chapter 4. p41: strategic entropy of a choice point, with error bars
# Model (illustrative counts, not real game statistics): how 1000
# professional games answered one standard corner position, before
# and after the AI era, over six named options. Strategic entropy is
# Shannon's H in bits; 2^H is the effective number of options. The
# Miller-Madow term (K - 1) / (2 N ln 2) corrects the plug-in bias;
# a seeded bootstrap (500 resamples) gives 95 % intervals. The last
# block shows how badly H is underestimated from a small sample.
import math
import random

SEED = 20261112
OPTIONS = ["A", "B", "C", "D", "E", "F"]
PRE = [620, 210, 90, 50, 20, 10]
POST = [300, 250, 200, 120, 80, 50]


def entropy(counts):
    n = math.fsum(counts)
    return -math.fsum(c / n * math.log2(c / n) for c in counts if c > 0)


def miller_madow(counts):
    n = math.fsum(counts)
    k = len([c for c in counts if c > 0])
    return entropy(counts) + (k - 1) / (2.0 * n * math.log(2.0))


def resample(rng, counts, n):
    pool = []
    for i, c in enumerate(counts):
        pool.extend([i] * c)
    out = [0] * len(counts)
    for _ in range(n):
        out[rng.choice(pool)] += 1
    return out


def boot_ci(rng, counts, reps=500):
    n = int(math.fsum(counts))
    hs = sorted(entropy(resample(rng, counts, n)) for _ in range(reps))
    return hs[int(0.025 * reps)], hs[int(0.975 * reps) - 1]


def js_div(a, b):
    na, nb = math.fsum(a), math.fsum(b)
    p = [x / na for x in a]
    q = [x / nb for x in b]
    m = [(x + y) / 2.0 for x, y in zip(p, q)]

    def kl(u, v):
        return math.fsum(x * math.log2(x / y) for x, y in zip(u, v)
                         if x > 0)
    return (kl(p, m) + kl(q, m)) / 2.0


if __name__ == "__main__":
    rng = random.Random(20261112)
    print("era    H bits  MM bits  2^H    95 % bootstrap CI  top share")
    for name, c in (("pre ", PRE), ("post", POST)):
        h = entropy(c)
        lo, hi = boot_ci(rng, c)
        print("{}   {:.3f}   {:.3f}   {:.2f}   [{:.3f}, {:.3f}]"
              "     {:.3f}".format(name, h, miller_madow(c), 2 ** h,
                                   lo, hi, max(c) / math.fsum(c)))
    print("maximum for 6 options: {:.3f} bits".format(math.log2(6)))
    print("Jensen-Shannon divergence pre vs post: {:.4f} bits".format(
        js_div(PRE, POST)))
    print("small samples from the post-AI law (300 draws each):")
    print("  n     mean plug-in H   mean Miller-Madow H")
    for n in (10, 30, 100, 1000):
        hp, hm = [], []
        for _ in range(300):
            s = resample(rng, POST, n)
            hp.append(entropy(s))
            hm.append(miller_madow(s))
        print("  {:4d}   {:.3f}            {:.3f}".format(
            n, math.fsum(hp) / 300, math.fsum(hm) / 300))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《围棋人工智能后人类决策质量跃迁、新颖着法率与棋经十三篇策略熵母典》第 4 章「策略熵:定式集中度、局部最优与从众锁定」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-230 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026230,
  author    = {Fan, Yafu},
  title     = {Post-AI Leaps in Human Go Decision Quality, Novel-Move Rates and Strategic Entropy in the Thirteen-Chapter Classic of Go},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论