关于「两种摩擦:换线成本要求滞后带,会变的市场要求记录褪色」,《业务线周度配额》给出的定量结论是什么?

结论:有换线成本时,吉廷斯指数不再有最优性,滞后带是一个简单有效的补丁:成本越高,最好的滞后带越宽,在示意模型里约为成本的三到五成。市场会变时,旧记录必须褪色:γ 取 0.97 时,汤普森抽样与贪心规则都把收益从理想值的七成多提高到近九成。到这里,规则里的「一次成交」还只是一个抽象的单位。收费站、代购、出版三条线的一次成交,价值相差很大,单位成本也不同;指数该怎样换算成钱?

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: 0.97

restless lines: rate redrawn w.p. 0.01 per period, T = 2000
rulegamma share-oracle
greedy1.000.7160
thompson1.000.7699
thompson0.990.8707
thompson0.970.8829
thompson0.900.8461
greedy0.970.8849
查看完整验算程序
# Chapter 5. p52: when idle lines keep changing, memory must fade.
# Model: 3 Bernoulli lines; each period, each line's rate is redrawn
# from Uniform(0, 1) with probability 0.01 (a market shift that happens
# whether or not the line is funded: a restless line, Whittle 1988).
# One unit per period, 2000 periods, 100 runs, no discounting.
# Rules: greedy and Thompson on full memory, and Thompson with
# forgetting gamma: every period all pseudo-counts are multiplied by
# gamma, so old evidence fades and an idle line's uncertainty regrows.
# Score = mean successes per period / oracle (who knows current rates).
import math
import random

N, T, RUNS, SHIFT = 3, 2000, 100, 0.01


def run(rule, gamma, world_seed, rng):
    world = random.Random(world_seed)
    p = [world.random() for _ in range(N)]
    s, f = [0.0] * N, [0.0] * N
    hits, best = 0, []
    for _ in range(T):
        for i in range(N):
            if world.random() < SHIFT:
                p[i] = world.random()
        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))
        hit = world.random() < p[i]
        best.append(max(p))
        for j in range(N):
            s[j] *= gamma
            f[j] *= gamma
        if hit:
            s[i] += 1
            hits += 1
        else:
            f[i] += 1
    return hits / math.fsum(best)


print("restless lines: rate redrawn w.p. %.2f per period, T = %d"
      % (SHIFT, T))
print("%-10s %6s %12s" % ("rule", "gamma", "share-oracle"))
for rule, gamma in (("greedy", 1.0), ("thompson", 1.0),
                    ("thompson", 0.99), ("thompson", 0.97),
                    ("thompson", 0.9), ("greedy", 0.97)):
    rng = random.Random(5)
    sc = [run(rule, gamma, 1000 + k, rng) for k in range(RUNS)]
    print("%-10s %6.2f %12.4f" % (rule, gamma, math.fsum(sc) / 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 5 章「两种摩擦:换线成本要求滞后带,会变的市场要求记录褪色」(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