关于「贝叶斯更新与信息增益选题」,《出生时刻误差与交界翻转》给出的定量结论是什么?

结论:期望信息增益给出了一个透明的选题规则:每次问能让后验熵平均下降最多的那道题。在示意题库里,单题增益最多 0.1564 比特,远小于先验熵 1.5219 比特;一个人的后验可以在问答中大幅摇摆,停在哪一题会改变结论。按一万人平均,贪心法在四道题时的判错率是 0.2808,随机顺序是 0.3648;问满十二道题,所有策略都落在约两成的地板上。答题者记错年份与顺从倾向会进一步削弱每道题:记错概率为 0.3 时,原来 7 道够用的题要 15 道。下一章讨论最危险的情形:题目其实毫无区分力,而「事后对上」仍然频繁出现。

数据口径:本书只分析排盘规则的结构、误差传播与概率,不做任何个人占断,不宣称任何体系有预测效力;书中参数为示意,全部计算由书中程序生成。 第一章说明,示意模型中凭记忆报告的时刻有 0.2782 恰好落在时辰分界上,不取整时只有 0.0083。

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: 0.1564, 1.5219

expected gain at the prior (bits), H(prior)1.5219
Q1 Q2 Q3 Q4 Q5 Q6
0.1141 0.1141 0.0768 0.0777 0.1564 0.0058
Q7 Q8 Q9 Q10 Q11 Q12
0.1564 0.0000 0.1510 0.1510 0.0633 0.0763
trace, true candidate Si
step item gain answer Chen Si Wu
1 Q70.1564 yes 0.1818 0.4545 0.3636
2 Q50.1431 no 0.0656 0.4098 0.5246
3 Q40.1246 yes 0.1008 0.6303 0.2689
4 Q20.1109 yes 0.0548 0.7991 0.1461
5 Q30.0601 yes 0.0459 0.6688 0.2854
6 Q120.0973 yes 0.0346 0.5043 0.4611
7 Q90.0615 yes 0.0678 0.3296 0.6026
8 Q100.0626 yes 0.0300 0.4371 0.5329
9 Q10.0142 yes 0.0673 0.4203 0.5124
10 Q110.0167 yes 0.1181 0.3974 0.4845
11 Q60.0032 yes 0.1336 0.3680 0.4984
12 Q80.0000 yes 0.1336 0.3680 0.4984
stopped after 12 items; top candidate Wu
查看完整验算程序
# Chapter 5. p51: choose the next question by expected information gain.
# Three candidates (Chen, Si, Wu) with the "morning 07-12" prior of p41.
# A bank of twelve abstract past-event items; row k gives P(yes | Chen),
# P(yes | Si), P(yes | Wu). These numbers are illustrative only: no
# published study shows that any such item discriminates between
# double hours, and Chapter 6 shows what happens when none does. The
# expected gain of an item is I = H(post) - E[H(post | answer)]. We
# list the gains at the prior, then trace one simulated person whose
# true candidate is Si, asking greedily without repeating an item and
# stopping when one candidate reaches 0.95. Fixed seed.
import math
import random

NAMES = ["Chen", "Si", "Wu"]
PRIOR = [0.4, 0.4, 0.2]
BANK = [(0.70, 0.30, 0.30), (0.30, 0.70, 0.30), (0.30, 0.30, 0.70),
        (0.60, 0.60, 0.20), (0.80, 0.50, 0.20), (0.55, 0.45, 0.50),
        (0.20, 0.50, 0.80), (0.50, 0.50, 0.50), (0.75, 0.25, 0.50),
        (0.25, 0.75, 0.50), (0.65, 0.35, 0.35), (0.35, 0.35, 0.75)]


def entropy(p):
    return -math.fsum(x * math.log2(x) for x in p if x > 0)


def update(post, item, yes):
    w = [p * (y if yes else 1 - y) for p, y in zip(post, item)]
    z = math.fsum(w)
    return [x / z for x in w], z


def gain(post, item):
    py_post, py = update(post, item, True)
    pn_post, pn = update(post, item, False)
    return entropy(post) - py * entropy(py_post) - pn * entropy(pn_post)


print("expected gain at the prior (bits), H(prior) = %.4f"
      % entropy(PRIOR))
gains = [gain(PRIOR, it) for it in BANK]
print("  " + " ".join("Q%-5d" % (k + 1) for k in range(6)))
print("  " + " ".join("%.4f" % g for g in gains[:6]))
print("  " + " ".join("Q%-5d" % (k + 1) for k in range(6, 12)))
print("  " + " ".join("%.4f" % g for g in gains[6:]))
rng = random.Random(20261113)
truth = 1
post, used = list(PRIOR), []
print("trace, true candidate Si")
print("step item  gain    answer   Chen    Si      Wu")
while max(post) < 0.95 and len(used) < len(BANK):
    k = max((gain(post, BANK[j]), -j) for j in range(len(BANK))
            if j not in used)[1] * -1
    g = gain(post, BANK[k])
    yes = rng.random() < BANK[k][truth]
    post, _ = update(post, BANK[k], yes)
    used.append(k)
    print("%3d  Q%-3d %7.4f  %-6s %7.4f %7.4f %7.4f"
          % (len(used), k + 1, g, "yes" if yes else "no", *post))
print("stopped after %d items; top candidate %s"
      % (len(used), NAMES[post.index(max(post))]))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《出生时刻记忆误差传播与鲁棒定盘、交界时辰盘面翻转概率与最少前事反推题数母典》第 5 章「贝叶斯更新与信息增益选题」(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-284 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026284,
  author    = {Fan, Yafu},
  title     = {Error Propagation from Remembered Birth Times and Robust Chart Fixing: Flip Probabilities at Double-Hour Boundaries and the Fewest Past-Event Questions},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论