关于「自适应选题:最大信息准则与题数的信息论下界」,《最少题数的信息论》给出的定量结论是什么?
结论:最大信息准则让下一题跟着当前估计走。模拟中自适应五题的误差 0.3866,略优于固定八题的 0.3969;按精度停止平均只需 4.33 题。信息论给出硬性下界:每道是非题至多 \(1 - H(e)\) 比特,回答错误率 0.2 时从十二个时段中选一至少要问 13 题。下一章把视野从一个参数扩展到两个:当回答里还混着「爱说是」的作答风格时,费希尔信息矩阵告诉我们该怎样出题。
数据口径:书中的价格、预算、题库与判别力参数均为示意;全部计算由书中程序生成;本书是非个人化经营建议;涉及出生时刻的部分只讨论估计理论与问卷设计,不评价任何体系的预测效度,本书不做任何个人占断。 第四章的模拟中,自适应五题的误差 0.3866 优于固定八题的 0.3969,按精度停止平均只需 4.33 题;信道容量给出任何问法都绕不过的题数下界。
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: 0.3866, 0.3969, 4.33
| design | mean items | RMSE | mean post SD | P(SD < 0.42) |
| fixed ladder 5 | 5.00 | 0.4524 | 0.4369 | 0.4655 |
| fixed ladder 8 | 8.00 | 0.3969 | 0.3801 | 0.9315 |
| adaptive 3 | 3.00 | 0.4742 | 0.4647 | 0.0000 |
| adaptive 5 | 5.00 | 0.3866 | 0.3759 | 0.9395 |
| adaptive stop | 4.33 | 0.4037 | 0.3991 | 1.0000 |
| respondents per design: 2,000; bank of 25 prices |
查看完整验算程序
# Chapter 4. p41: adaptive selection by maximum Fisher information
# versus fixed price ladders. Fixed-seed Monte Carlo, illustrative.
# 题库:25 个价格门槛,从 500 到 24,000 按对数等距。每一步选在当前
# 后验均值处费希尔信息最大的未用题;后验标准差低于 0.42 或问满上限即停。
# 对照:固定五题阶梯(p12)、固定八题阶梯。先验与 a 同 p31。
import math
import random
from statistics import NormalDist
A, MU, SD, N = 2.5, math.log(3000), 0.8, 2000
BANK = [math.log(500) + i * math.log(48) / 24 for i in range(25)]
FIX5 = [math.log(x) for x in (1000, 2000, 3000, 5000, 8000)]
FIX8 = [math.log(x) for x in (700, 1100, 1700, 2500, 3600, 5300, 7800,
11500)]
GRID = [MU - 4 * SD + 8 * SD * i / 240 for i in range(241)]
PRIOR = [NormalDist(MU, SD).pdf(t) for t in GRID]
def p_yes(theta, b):
return 1.0 / (1.0 + math.exp(-A * (theta - b)))
def update(w, b, y):
out = [v * (p_yes(t, b) if y else 1.0 - p_yes(t, b))
for t, v in zip(GRID, w)]
z = math.fsum(out)
return [v / z for v in out]
def moments(w):
m = math.fsum(t * v for t, v in zip(GRID, w))
s = math.sqrt(math.fsum((t - m) ** 2 * v for t, v in zip(GRID, w)))
return m, s
def run(th, rng, items=None, stop_sd=None, cap=8):
z = math.fsum(PRIOR)
w = [v / z for v in PRIOR]
used, m, s = [], MU, SD
while len(used) < cap:
if items is not None:
b = items[len(used)]
else:
if stop_sd is not None and s < stop_sd:
break
b = min((x for x in BANK if x not in used),
key=lambda x: abs(x - m))
used.append(b)
w = update(w, b, rng.random() < p_yes(th, b))
m, s = moments(w)
return m, s, len(used)
DESIGNS = (("fixed ladder 5", FIX5, None, 5),
("fixed ladder 8", FIX8, None, 8),
("adaptive 3", None, None, 3),
("adaptive 5", None, None, 5),
("adaptive stop", None, 0.42, 8))
print("design mean items RMSE mean post SD P(SD < 0.42)")
for name, items, stop, cap in DESIGNS:
rng = random.Random(20261113)
sq, sds, ns, ok = [], [], [], 0
for _ in range(N):
th = rng.gauss(MU, SD)
m, s, k = run(th, rng, items, stop, cap)
sq.append((m - th) ** 2)
sds.append(s)
ns.append(k)
ok += s < 0.42
print("%-15s %-11.2f %.4f %.4f %.4f"
% (name, math.fsum(ns) / N, math.sqrt(math.fsum(sq) / N),
math.fsum(sds) / N, ok / N))
print("respondents per design: {:,}; bank of {} prices".format(N,
len(BANK)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《费希尔信息矩阵、克拉美罗方差下界与预算及出生时刻探询最少题数母典》第 4 章「自适应选题:最大信息准则与题数的信息论下界」(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-235 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026235,
author = {Fan, Yafu},
title = {Fisher Information Matrices and the Cramér-Rao Bound: the Fewest Questions Needed to Estimate a Budget or a Birth Time},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}