关于「时间上的断点:二〇一六年三月的回归不连续设计」,《围棋 AI 后的人类决策质量》给出的定量结论是什么?
结论:第二章在模拟数据上演示时间断点回归:真实跳变为 0.30 时,12 个月带宽的检出率只有 0.685,而一个只有斜坡的世界在 36 个月带宽下有 0.258 的机会被判为显著。
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。
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: 0.258, 0.685, 12, 36
| part 1: placebo cutoffs inside 2011 to 2015, h = 12 | ||||
| placebo cutoffs 51, |z| > 1.96 in 4 (0.078) | ||||
| part 2: 400 redraws per world | ||||
| h | A mean | A reject | B mean | B reject |
| 6 | +0.298 | 0.405 | -0.005 | 0.083 |
| 12 | +0.301 | 0.685 | -0.009 | 0.070 |
| 24 | +0.231 | 0.757 | -0.063 | 0.102 |
| 36 | +0.208 | 0.838 | -0.097 | 0.258 |
查看完整验算程序
# Chapter 2. p22: placebo cutoffs and repeated simulation of p21
# Same simulated worlds as p21 (World A: 0.30 jump in March 2016 and a
# 0.02 a month ramp from January 2017; World B: ramp only). Part 1
# moves the cutoff to every pre-AI month whose window stays inside
# 2010 to February 2016 and counts how often a false jump is "found".
# Part 2 redraws each world 400 times and reports the mean estimate
# and the share of runs with |z| > 1.96 at four bandwidths. All
# parameters are illustrative; the functions are copied from p21.
import math
import random
SEED = 20261112
MONTHS = 144
CUT = 74
RAMP0 = 84
REPS = 400
def series(rng, jump, ramp):
ys = []
for t in range(MONTHS):
parts = [98.0, 0.002 * (t - CUT), rng.gauss(0.0, 0.15)]
if t >= CUT:
parts.append(jump)
if t >= RAMP0:
parts.append(ramp * (t - RAMP0 + 1))
ys.append(math.fsum(parts))
return ys
def solve(a, b):
n = len(b)
m = [row[:] + [b[i]] + [1.0 if j == i else 0.0 for j in range(n)]
for i, row in enumerate(a)]
for c in range(n):
p = max(range(c, n), key=lambda r: abs(m[r][c]))
m[c], m[p] = m[p], m[c]
piv = m[c][c]
m[c] = [v / piv for v in m[c]]
for r in range(n):
if r != c:
f = m[r][c]
m[r] = [v - f * w for v, w in zip(m[r], m[c])]
return [m[i][n] for i in range(n)], [m[i][n + 1:] for i in range(n)]
def rd_jump(ys, cut, h):
xs, yy = [], []
for t, y in enumerate(ys):
if cut - h <= t < cut + h:
d = 1.0 if t >= cut else 0.0
r = float(t - cut)
xs.append([1.0, d, r, d * r])
yy.append(y)
xtx = [[math.fsum(x[i] * x[j] for x in xs) for j in range(4)]
for i in range(4)]
xty = [math.fsum(x[i] * y for x, y in zip(xs, yy)) for i in range(4)]
beta, inv = solve(xtx, xty)
res = [y - math.fsum(b * v for b, v in zip(beta, x))
for x, y in zip(xs, yy)]
s2 = math.fsum(r * r for r in res) / (len(yy) - 4)
return beta[1], math.sqrt(s2 * inv[1][1])
if __name__ == "__main__":
rng = random.Random(20261112)
world_a = series(rng, 0.30, 0.02)
fakes = list(range(12, CUT - 12 + 1))
hits = 0
for c in fakes:
j, s = rd_jump(world_a, c, 12)
if abs(j / s) > 1.96:
hits += 1
print("part 1: placebo cutoffs inside 2011 to 2015, h = 12")
print(" placebo cutoffs {}, |z| > 1.96 in {} ({:.3f})".format(
len(fakes), hits, hits / len(fakes)))
print("part 2: {} redraws per world".format(REPS))
print(" h A mean A reject B mean B reject")
for h in (6, 12, 24, 36):
rows = []
for jump in (0.30, 0.0):
est, rej = [], 0
for _ in range(REPS):
j, s = rd_jump(series(rng, jump, 0.02), CUT, h)
est.append(j)
if abs(j / s) > 1.96:
rej += 1
rows.append((math.fsum(est) / REPS, rej / REPS))
print(" {:2d} {:+.3f} {:.3f} {:+.3f} {:.3f}".format(
h, rows[0][0], rows[0][1], rows[1][0], rows[1][1]))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 2 章「时间上的断点:二〇一六年三月的回归不连续设计」(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}
}