关于「锁死:按上周回报全押为何会长期困在次优业务线」,《业务线周度配额》给出的定量结论是什么?

结论:第一章量化锁死,并显示视界越长代价越大:视界 500 期时贪心每期后悔停在 0.0563,汤普森抽样降到 0.0216。

数据口径:核对吉廷斯 1979 年的论文与后续文献以后,本书修正了立项蓝图:吉廷斯指数的最优性只在单线经营、闲置线不变、换线无成本、几何贴现时成立,周度拆分预算、换线成本与会变的市场都在定理之外;锁死是一个概率说法,两条示意线上按累计回报全押有 0.2133 的重复在下半年完全押错,但在每单位贴现 0.9 的短视界里贪心相对指数只损失 0.0050 到 0.0166。 书中参数为示意,毛利与成本用记账点数而非价格;全部计算由书中程序生成。

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: 0.0216, 0.0563, 500

NTgrd-wrongts-wrong grd-regretts-regret
2200.20950.17350.05060.0698
21000.18000.09400.02730.0266
25000.16670.04000.01870.0077
4200.41600.34900.10840.1603
41000.36200.15500.06990.0690
45000.33330.05670.05630.0216
查看完整验算程序
# Chapter 1. p13: the longer the horizon, the more greedy lock-in costs.
# Model: N Bernoulli lines with rates drawn from Uniform(0, 1), one unit
# per period, undiscounted horizon T periods. Greedy plays the highest
# posterior mean under Beta(1, 1); Thompson plays the highest posterior
# draw. Wrong-line = the line played most often is not the best line.
# Regret per period = (T max p - successes) / T, averaged over runs.
# Common random numbers as in p12. Illustrative; fixed-seed Monte Carlo.
import math
import random


def play(rule, p, u, horizon, rng):
    n = len(p)
    s, f, pulls = [0] * n, [0] * n, [0] * n
    for _ in range(horizon):
        if rule == "greedy":
            sc = [(1 + s[i]) / (2 + s[i] + f[i]) for i in range(n)]
        else:
            sc = [rng.betavariate(1 + s[i], 1 + f[i]) for i in range(n)]
        i = max(range(n), key=lambda j: (sc[j], -j))
        pulls[i] += 1
        if u[i][s[i] + f[i]] < p[i]:
            s[i] += 1
        else:
            f[i] += 1
    top = max(range(n), key=lambda j: (pulls[j], -j))
    best = max(range(n), key=lambda j: (p[j], -j))
    hits = 0
    for x in s:
        hits += x
    return top != best, (horizon * p[best] - hits) / horizon


print("%3s %5s %10s %10s %10s %10s" % ("N", "T", "grd-wrong", "ts-wrong",
                                       "grd-regret", "ts-regret"))
for n in (2, 4):
    for horizon, runs in ((20, 2000), (100, 1000), (500, 300)):
        world = random.Random(20261113)
        rng = random.Random(11)
        wrong = {"greedy": 0, "thompson": 0}
        reg = {"greedy": [], "thompson": []}
        for _ in range(runs):
            p = [world.random() for _ in range(n)]
            u = [[world.random() for _ in range(horizon)]
                 for _ in range(n)]
            for rule in ("greedy", "thompson"):
                w, r = play(rule, p, u, horizon, rng)
                wrong[rule] += w
                reg[rule].append(r)
        print("%3d %5d %10.4f %10.4f %10.4f %10.4f" % (
            n, horizon, wrong["greedy"] / runs, wrong["thompson"] / runs,
            math.fsum(reg["greedy"]) / runs,
            math.fsum(reg["thompson"]) / runs))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《多业务线算力与预算动态裁决、吉廷斯指数探索与收费站代购出版线周度配额母典》第 1 章「锁死:按上周回报全押为何会长期困在次优业务线」(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-264 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026264,
  author    = {Fan, Yafu},
  title     = {Dynamic Allocation of Compute and Budget across Business Lines: Gittins-Index Exploration and Weekly Quotas for Tollbooth, Buying and Publishing Lines},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论