关于「阈值漂移:折扣贝塔、滑动窗口与累积和报警」,《条件参数化规则蒸馏与在线校准》给出的定量结论是什么?

结论:第五章比较折扣、滑动窗口与累积和报警,滑动窗口在台阶变化下合计遗憾 33.0。

数据口径:书中参数为示意,全部计算由书中程序生成。 本书是非个人化经营分析,不构成投资、法律或医疗意见,不对应任何真实客户或成交。

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: 33.0

sessions 12,000, change at 5,000, runs 30
policyregret beforeregret aftertotalp90 total
fixed0.084.084.084.0
stationary10.160.770.885.4
discount21.229.050.257.9
window11.221.833.049.4
effective memory of the discount 0.998: 500 sessions
查看完整验算程序
# Chapter 5. p51: thresholds drift when traffic changes. Two copy tiers;
# at session 5,000 of 12,000 the traffic mix changes and their true pay
# rates swap order: tier 0 goes 0.045 -> 0.030, tier 1 goes
# 0.032 -> 0.042 (illustrative). 30 runs per policy:
#   fixed     : tier 0 forever (the threshold set before the change)
#   stationary: Thompson sampling with all history weighted equally
#   discount  : Thompson with counts multiplied by 0.998 each session
#   window    : Thompson on the last 1,500 sessions of each tier only
# Regret is measured against the best tier of each period.
import collections
import math
import random

T, CHANGE, RUNS = 12000, 5000, 30
BEFORE, AFTER = (0.045, 0.032), (0.030, 0.042)


def run(policy, rng):
    a = [0.0, 0.0]
    b = [0.0, 0.0]
    hist = [collections.deque(), collections.deque()]
    lost_pre, lost_post = [], []
    for t in range(T):
        p = BEFORE if t < CHANGE else AFTER
        if policy == "fixed":
            arm = 0
        else:
            draws = [rng.betavariate(1.0 + a[i], 1.0 + b[i])
                     for i in range(2)]
            arm = 0 if draws[0] >= draws[1] else 1
        r = 1 if rng.random() < p[arm] else 0
        if policy == "discount":
            for i in range(2):
                a[i] *= 0.998
                b[i] *= 0.998
        a[arm] += r
        b[arm] += 1 - r
        if policy == "window":
            hist[arm].append(r)
            if len(hist[arm]) > 1500:
                old = hist[arm].popleft()
                a[arm] -= old
                b[arm] -= 1 - old
        gap = max(p) - p[arm]
        (lost_pre if t < CHANGE else lost_post).append(gap)
    return math.fsum(lost_pre), math.fsum(lost_post)


def main():
    seeds = {"fixed": random.Random(5101),
             "stationary": random.Random(5102),
             "discount": random.Random(5103),
             "window": random.Random(5104)}
    print("sessions {:,}, change at {:,}, runs {}".format(T, CHANGE, RUNS))
    print(" policy       regret before  regret after  total   p90 total")
    for pol in ("fixed", "stationary", "discount", "window"):
        pre, post, tot = [], [], []
        for _ in range(RUNS):
            x, y = run(pol, seeds[pol])
            pre.append(x)
            post.append(y)
            tot.append(x + y)
        tot.sort()
        print(" {:<11}  {:>10.1f}    {:>10.1f}   {:>6.1f}   {:>6.1f}".format(
            pol, math.fsum(pre) / RUNS, math.fsum(post) / RUNS,
            math.fsum(tot) / RUNS, tot[26]))
    print("effective memory of the discount 0.998: {:.0f} sessions".format(
        1.0 / (1.0 - 0.998)))


if __name__ == "__main__":
    main()

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《高阶知识蒸馏进化、条件参数化编译与实盘转化反馈校准母典》第 5 章「阈值漂移:折扣贝塔、滑动窗口与累积和报警」(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-377 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026377,
  author    = {Fan, Yafu},
  title     = {Next-Generation Knowledge Distillation, Context-Parametric Compilation and Live-Conversion Feedback Calibration},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论