关于「断点上的跃迁:KataGo 量尺下的决策质量指数」,《围棋 AI 后的人类决策质量》给出的定量结论是什么?

结论:第一章把论文表 1 换算成效应量:AI 后普通着法的决策质量指数高出 0.59754 个百分点,新着高出 1.11258 个点,新着原有的代价缩小了 0.848;并说明匹配率超过一半时中位数汇总会碰到上限。

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: 0.59754, 0.848, 1.11258

term est se z95 % CI
after AI (b1)0.59754 0.01601 37.3 [0.5662, 0.6289]
novel (b2)-0.60770 0.01219 -49.9 [-0.6316, -0.5838]
after x novel (b3)0.51504 0.02147 24.0 [0.4730, 0.5571]
model 2 b30.51498, se 0.02132: same to 4 decimals
cell means relative to pre-AI, non-novel moves (DQI points)
pre-AI non-novel +0.00000
pre-AI novel-0.60770
post-AI non-novel +0.59754
post-AI novel +0.50488
AI gain, non-novel moves0.59754
AI gain, novel moves1.11258 (1.862 x)
novelty penalty before-0.60770
novelty penalty after-0.09266
worst-case se of b1 + b30.03748, lower 95 % bound 1.0391
share of b2 penalty removed by b30.848
查看完整验算程序
# Chapter 1. p11: what the published post-AI coefficients imply
# Source: Shin, Kim, van Opheusden and Griffiths (2023, PNAS), Table 1.
# Dependent variable: Decision Quality Index (DQI) of each move, in
# percentage points of win rate, DQI = 100 - (AI win rate - human win
# rate), estimated with KataGo (10000 simulated game patterns per
# decision); 5857513 moves, first 60 moves of each game, 1950 to 2021,
# player and move-number fixed effects, standard errors clustered by
# player. Only the coefficients and standard errors below are taken
# from the paper; every derived number is arithmetic on them.
import math

N_OBS = 5857513
PLAYOUTS = 10000
MAX_MOVE = 602
Z = 1.959964

# name: (model 1 estimate, se, model 2 estimate, se)
TABLE = [
    ("after AI (b1)", 0.59754, 0.01601, None, None),
    ("novel (b2)", -0.60770, 0.01219, -0.60777, 0.01218),
    ("after x novel (b3)", 0.51504, 0.02147, 0.51498, 0.02132),
]


def ci(b, se):
    return b - Z * se, b + Z * se


if __name__ == "__main__":
    print("term                 est       se      z      95 % CI")
    for name, b, se, b2, se2 in TABLE:
        lo, hi = ci(b, se)
        print("{:<18} {:9.5f} {:8.5f} {:6.1f}  [{:.4f}, {:.4f}]".format(
            name, b, se, b / se, lo, hi))
    b1, s1 = TABLE[0][1], TABLE[0][2]
    b2, s2 = TABLE[1][1], TABLE[1][2]
    b3, s3 = TABLE[2][1], TABLE[2][2]
    print("model 2 b3 = {:.5f}, se {:.5f}: same to 4 decimals".format(
        TABLE[2][3], TABLE[2][4]))
    print("cell means relative to pre-AI, non-novel moves (DQI points)")
    cells = [("pre-AI  non-novel", 0.0), ("pre-AI  novel", b2),
             ("post-AI non-novel", b1),
             ("post-AI novel", math.fsum([b1, b2, b3]))]
    for name, v in cells:
        print("  {:<18} {:+.5f}".format(name, v))
    gain_novel = math.fsum([b1, b3])
    print("AI gain, non-novel moves: {:.5f}".format(b1))
    print("AI gain, novel moves:     {:.5f}  ({:.3f} x)".format(
        gain_novel, gain_novel / b1))
    print("novelty penalty before:   {:.5f}".format(b2))
    print("novelty penalty after:    {:.5f}".format(math.fsum([b2, b3])))
    worst = math.fsum([s1, s3])
    print("worst-case se of b1 + b3: {:.5f}, lower 95 % bound {:.4f}".format(
        worst, gain_novel - Z * worst))
    print("share of b2 penalty removed by b3: {:.3f}".format(-b3 / b2))

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 1 章「断点上的跃迁:KataGo 量尺下的决策质量指数」(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