关于「校准法:把一条业务线与永久安全收益相比,二分求出吉廷斯指数」,《业务线周度配额》给出的定量结论是什么?

结论:第二章用校准法与重启法计算吉廷斯指数,从未试过的线在贴现 0.9 时指数为 0.7029。

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: 0.7029, 0.9

bisection for Beta(1, 1), beta = 0.9 (first 6 steps)
steplamretirecontinueverdict
10.750007.500007.34844retire
20.625006.250006.53388explore
30.687506.875006.92889explore
40.718757.187507.13536retire
50.703127.031257.03043retire
60.695316.953136.97948explore
index of Beta(1, 1), the untried line, by discount factor
beta 0.50depth40index 0.5590bonus +0.0590
beta 0.70depth40index 0.6046bonus +0.1046
beta 0.90depth40index 0.7029bonus +0.2029
beta 0.95depth80index 0.7614bonus +0.2614
beta 0.99depth 160index 0.8698bonus +0.3698

核心公式

\[V(a,b)=\max\Big(\frac{\lambda}{1-\beta},\; p\,[1+\beta V(a+1,b)]+(1-p)\,\beta V(a,b+1)\Big),\quad p=\frac{a}{a+b}.\]
查看完整验算程序
# Chapter 2. p21: the Gittins index by calibration against a safe arm.
# Model: a line in state Beta(a, b) faces a safe arm that pays lam per
# period for ever. Value with retirement allowed:
#   V(a, b) = max( lam / (1 - beta),
#                  p (1 + beta V(a + 1, b)) + (1 - p) beta V(a, b + 1) )
# with p = a / (a + b). The index is the lam at which retiring now and
# continuing are equally good (Gittins and Jones 1974; Whittle 1980).
# Truncation: after depth steps learning stops and the line is worth
# max(lam, p) / (1 - beta). Bisection on lam to 1e-6.
import math


def cont_value(a, b, beta, lam, depth):
    r = lam / (1.0 - beta)
    n = a + b + depth
    v = [max(r, (a + k) / n / (1.0 - beta)) for k in range(depth + 1)]
    for d in range(depth - 1, 0, -1):
        n = a + b + d
        nv = []
        for k in range(d + 1):
            p = (a + k) / n
            c = math.fsum((p * (1.0 + beta * v[k + 1]),
                           (1.0 - p) * beta * v[k]))
            nv.append(c if c > r else r)
        v = nv
    p = a / (a + b)
    return math.fsum((p * (1.0 + beta * v[1]), (1.0 - p) * beta * v[0]))


def gittins(a, b, beta, depth=40, tol=1e-6):
    lo, hi = a / (a + b), 1.0
    while hi - lo > tol:
        lam = 0.5 * (lo + hi)
        r = lam / (1.0 - beta)
        n = a + b + depth
        v = [max(r, (a + k) / n / (1.0 - beta)) for k in range(depth + 1)]
        for d in range(depth - 1, -1, -1):
            n = a + b + d
            nv = []
            for k in range(d + 1):
                p = (a + k) / n
                c = math.fsum((p * (1.0 + beta * v[k + 1]),
                               (1.0 - p) * beta * v[k]))
                nv.append(c if c > r else r)
            v = nv
        if v[0] > r + 1e-12:
            lo = lam
        else:
            hi = lam
    return 0.5 * (lo + hi)


print("bisection for Beta(1, 1), beta = 0.9 (first 6 steps)")
print("%4s %9s %11s %11s %9s" % ("step", "lam", "retire", "continue",
                                 "verdict"))
lo, hi = 0.5, 1.0
for step in range(1, 7):
    lam = 0.5 * (lo + hi)
    ret = lam / (1 - 0.9)
    con = cont_value(1, 1, 0.9, lam, 40)
    better = con > ret
    print("%4d %9.5f %11.5f %11.5f %9s" % (
        step, lam, ret, con, "explore" if better else "retire"))
    if better:
        lo = lam
    else:
        hi = lam
print("index of Beta(1, 1), the untried line, by discount factor")
for beta, depth in ((0.5, 40), (0.7, 40), (0.9, 40), (0.95, 80),
                    (0.99, 160)):
    g = gittins(1, 1, beta, depth)
    print("  beta %.2f  depth %3d  index %.4f  bonus %+.4f" % (
        beta, depth, g, g - 0.5))

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 2 章「校准法:把一条业务线与永久安全收益相比,二分求出吉廷斯指数」(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