关于「成本曲线:准确率更高的机器人为什么反而更贵」,《三级灯号切分点》给出的定量结论是什么?
结论:第三章用成本曲线比较两个机器人,准确率较低但更会拒判的那个在成本比高于 4.14 时更省钱。
数据口径:本书用纯标准库程序把命题写成可计算的模型,并修正蓝图三处:带拒识通道的最优识别早见于 Chow 1957 年的论文,1970 年的论文给出误差与拒判的一般关系;规则只对校准过的概率成立,把成本比算出的切分点直接套在过度自信机器人的自报分数上,每百段对话损失 265.4 元,等渗回归校准后 239.8 元;私域客服有「机器人起草、真人过目」的第三种动作,需要六格损失矩阵与两条切分点,示意参数下答对概率低于 0.4000 亮红灯、不低于 0.9500 亮绿灯。 书中参数为示意,对话全部虚构;全部计算由书中程序生成;程序不读写任何文件,不连接任何客服系统;本书是非个人化经营建议,任何线上灯号的改动须由主事者人工复核放行。
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: 4.14
| bot A accuracy 0.897 | AUC 0.842 | |||
| bot B accuracy 0.819 | AUC 0.957 | |||
| LW | LW/CR | cost A | cost B | cheaper |
| 10 | 1.25 | 102.8 | 169.1 | A |
| 16 | 2.00 | 158.0 | 211.0 | A |
| 24 | 3.00 | 217.1 | 244.4 | A |
| 32 | 4.00 | 264.4 | 267.6 | A |
| 50 | 6.25 | 343.7 | 303.3 | B |
| 80 | 10.00 | 430.5 | 340.9 | B |
| 128 | 16.00 | 514.8 | 378.5 | B |
| 256 | 32.00 | 623.4 | 433.4 | B |
| break-even wrong-answer loss LW = 33.1 (LW/CR = 4.14) |
查看完整验算程序
# Chapter 3. p31: the more accurate bot is not always the cheaper one.
# Model (illustrative): two calibrated bots, p = 1 / (1 + e^-z).
# Bot A: z ~ N(3.0, 1.6), the p11 bot. Bot B: z ~ N(4.0, 4.0), less
# accurate overall but with more of its mass at the extremes. Each
# uses its own Chow cut 1 - CR / LW with CR = 8; the table gives
# yuan per 100 chats as LW moves, i.e. a cost curve in the spirit of
# Drummond and Holte (2006). AUC is the chance that a chat the bot
# gets right outranks one it gets wrong (exact, by quadrature).
import math
CR = 8.0
def nodes(mu, sd, n=4001):
zs = [mu + sd * (-8.0 + 16.0 * i / (n - 1)) for i in range(n)]
ws = [math.exp(-0.5 * ((z - mu) / sd) ** 2) for z in zs]
tot = math.fsum(ws)
return [(1.0 / (1.0 + math.exp(-z)), w / tot) for z, w in zip(zs, ws)]
def auc(pw):
pr = math.fsum(w * p for p, w in pw)
below, terms = [], []
for p, w in pw: # nodes are sorted by p already
terms.append(w * p * math.fsum(below))
below.append(w * (1 - p))
return math.fsum(terms) / (pr * (1 - pr))
def cost(pw, lw):
return 100 * math.fsum(w * min((1 - p) * lw, CR) for p, w in pw)
A, B = nodes(3.0, 1.6), nodes(4.0, 4.0)
print("bot A accuracy %.3f AUC %.3f" % (
math.fsum(w * p for p, w in A), auc(A)))
print("bot B accuracy %.3f AUC %.3f" % (
math.fsum(w * p for p, w in B), auc(B)))
print(" LW LW/CR cost A cost B cheaper")
for lw in (10.0, 16.0, 24.0, 32.0, 50.0, 80.0, 128.0, 256.0):
ca, cb = cost(A, lw), cost(B, lw)
print("%6.0f %5.2f %7.1f %7.1f %7s"
% (lw, lw / CR, ca, cb, "A" if ca <= cb else "B"))
lo, hi = 10.0, 128.0
for _ in range(60):
mid = (lo + hi) / 2
if cost(A, mid) <= cost(B, mid):
lo = mid
else:
hi = mid
print("break-even wrong-answer loss LW = %.1f (LW/CR = %.2f)"
% (lo, lo / CR))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《私域客服人机交接三级灯号精算、误判损失矩阵与最优置信度切分点母典》第 3 章「成本曲线:准确率更高的机器人为什么反而更贵」(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-266 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026266,
author = {Fan, Yafu},
title = {Three Lights for Handing Over between Bot and Human in Private Customer Service: Misclassification Loss Matrices, Chow's Reject Rule and Optimal Cut Points on Calibrated Confidence},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}