关于「条件参数化编译:把决策表压成决策列表与有序决策图」,《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:第二章把 72 格决策表编译成决策列表与有序决策图,变量次序不同,节点数从 6 个到 23 个不等。

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: 23, 72

orders 120; internal nodes min 6 median 12.0 max 23
best three orders (nodes, mean tests):
visit>src>dev>slot>tier6 2.702
visit>src>slot>tier>dev6 2.702
visit>src>slot>dev>tier6 2.702
worst three orders:
tier>dev>slot>visit>src23 4.052
tier>slot>dev>visit>src23 4.052
slot>tier>dev>visit>src23 4.552
fewest mean tests: visit>src>dev>slot>tier with 6 nodes, 2.702 tests
flat table rows 72; every diagram answers in <5 tests
查看完整验算程序
# Chapter 2. p22: the same conditioned table as a reduced ordered
# decision diagram (Bryant 1986, generalised to multi-valued tests).
# For each of the 5! = 120 attribute orders we build the reduced diagram
# (merge identical sub-diagrams, skip tests whose branches all agree)
# and count internal nodes and the traffic-weighted number of tests per
# query. Table and weights are the illustrative ones of p21.
import itertools
import math
import statistics

ATTRS = [("tier", ("low", "mid", "high"), (0.50, 0.35, 0.15)),
         ("src", ("search", "social", "direct"), (0.45, 0.35, 0.20)),
         ("dev", ("mobile", "desktop"), (0.70, 0.30)),
         ("visit", ("first", "return"), (0.65, 0.35)),
         ("slot", ("day", "night"), (0.60, 0.40))]


def label(c):
    tier, src, dev, visit, slot = c
    if visit == "return":
        return "S" if tier == "high" else "B"
    if src == "social":
        return "B" if dev == "mobile" else "S"
    if src == "search":
        return "S" if tier == "low" else "D"
    return "D" if (slot == "night" and tier != "low") else "S"


def build(order):
    unique = {}

    def node(level, partial):
        if level == len(order):
            c = tuple(partial[i] for i in range(len(ATTRS)))
            return ("leaf", label(c))
        a = order[level]
        kids = []
        for v in ATTRS[a][1]:
            partial[a] = v
            kids.append(node(level + 1, partial))
        del partial[a]
        if all(k == kids[0] for k in kids):
            return kids[0]
        key = (a, tuple(kids))
        if key not in unique:
            unique[key] = ("n", len(unique))
        return unique[key]

    root = node(0, {})
    return root, unique


def tests(order, root, unique):
    back = {v: k for k, v in unique.items()}
    terms = []
    for c in itertools.product(*[a[1] for a in ATTRS]):
        w = 1.0
        for (_, vals, ws), v in zip(ATTRS, c):
            w *= ws[vals.index(v)]
        n, steps = root, 0
        while n[0] == "n":
            a, kids = back[n]
            n = kids[ATTRS[a][1].index(c[a])]
            steps += 1
        terms.append(w * steps)
    return math.fsum(terms)


def main():
    rows = []
    for order in itertools.permutations(range(len(ATTRS))):
        root, unique = build(list(order))
        rows.append((len(unique), tests(order, root, unique), order))
    sizes = [r[0] for r in rows]
    print("orders 120; internal nodes min {} median {} max {}".format(
        min(sizes), statistics.median(sizes), max(sizes)))
    rows.sort(key=lambda r: (r[0], r[1], r[2]))
    name = lambda o: ">".join(ATTRS[i][0] for i in o)
    print("best three orders (nodes, mean tests):")
    for n, t, o in rows[:3]:
        print("  {:<30} {:>3} {:.3f}".format(name(o), n, t))
    print("worst three orders:")
    for n, t, o in rows[-3:]:
        print("  {:<30} {:>3} {:.3f}".format(name(o), n, t))
    fast = min(rows, key=lambda r: (r[1], r[0], r[2]))
    print("fewest mean tests: {} with {} nodes, {:.3f} tests".format(
        name(fast[2]), fast[0], fast[1]))
    print("flat table rows 72; every diagram answers in <= 5 tests")


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 2 章「条件参数化编译:把决策表压成决策列表与有序决策图」(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