关于「反事实遗憾最小化:从三牌暗押到卡组选择中的必胜坏牌组」,《卡牌费用回归与循环克制》给出的定量结论是什么?

结论:第五章实现反事实遗憾最小化,在三牌暗押上逼近理论值 −0.0556,并在一千轮内把 0.808 的份额推给建立在被低估卡上的坏牌组。

数据口径:本书用一款自拟的示意卡牌游戏「衡枢」把这些问题写成可以运行的算术。 书中卡池、胜率与价格参数均为示意;全部计算由书中程序生成;本书不是法律、财务或投资建议,也不推荐任何购买。

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: 0.808

before patch: average strategy A C Cb M T S
1000 iterations0.065 0.126 0.000 0.001 0.000 0.808
10000 iterations0.006 0.013 0.000 0.000 0.000 0.981
200000 iterations0.000 0.001 0.000 0.000 0.000 0.999
top deck Surge share 0.999, worst matchup 0.54: BROKEN
after patch: average strategy A C Cb M T S
1000 iterations0.325 0.271 0.092 0.019 0.121 0.172
10000 iterations0.411 0.298 0.214 0.038 0.012 0.026
200000 iterations0.440 0.333 0.215 0.009 0.001 0.001
top deck Aggro share 0.440, worst matchup 0.40: no flag
查看完整验算程序
# Chapter 5. p53: Sampled regret matching flags a must-pick broken deck.
# Five decks of p42 plus "Surge", built around the under-costed card
# Y07 of p11 (all win rates illustrative). Each iteration samples one
# opponent deck from the current strategy and one game result for every
# candidate deck (seed 20261112), then updates regrets: Monte Carlo CFR
# on the deck-choice move. Rule: a deck holding more than half of the
# average strategy is a broken-deck candidate; we confirm exactly that
# it wins at least half of every matchup. The patch moves Y07 to cost 4
# and lowers each Surge matchup by 0.07 (assumed; Chapter 6 estimates it).
import math
import random

D = ["Aggro", "Control", "Combo", "Midrange", "Tempo", "Surge"]
UP = {(0, 1): 40, (0, 2): 65, (0, 3): 60, (0, 4): 62,
      (1, 2): 30, (1, 3): 42, (1, 4): 45,
      (2, 3): 44, (2, 4): 46, (3, 4): 53}
SURGE = [0.56, 0.54, 0.58, 0.60, 0.63]
n = len(D)


def matrix(shift):
    W = [[0.5] * n for _ in range(n)]
    for (i, j), w in UP.items():
        W[i][j], W[j][i] = w / 100, 1 - w / 100
    for j, w in enumerate(SURGE):
        W[5][j] = w - shift
        W[j][5] = 1 - W[5][j]
    return W


def run(W, iters, rng):
    reg, tot = [0.0] * n, [0.0] * n
    marks = {}
    for t in range(1, iters + 1):
        pos = [max(r, 0.0) for r in reg]
        z = math.fsum(pos)
        x = [p / z for p in pos] if z > 0 else [1 / n] * n
        for i in range(n):
            tot[i] += x[i]
        j = rng.choices(range(n), weights=x)[0]
        o = [1.0 if rng.random() < W[i][j] else 0.0 for i in range(n)]
        ox = math.fsum(a * b for a, b in zip(x, o))
        for i in range(n):
            reg[i] += o[i] - ox
        if t in (1000, 10000, 200000):
            z = math.fsum(tot)
            marks[t] = [v / z for v in tot]
    return marks


rng = random.Random(20261112)
for label, shift in (("before patch", 0.0), ("after patch", 0.07)):
    W = matrix(shift)
    marks = run(W, 200000, rng)
    print(label + ": average strategy  A     C     Cb    M     T     S")
    for t, avg in marks.items():
        print("  {:6d} iterations    ".format(t)
              + " ".join("{:.3f}".format(v) for v in avg))
    avg = marks[200000]
    top = max(range(n), key=lambda i: avg[i])
    worst = min(W[top][j] for j in range(n) if j != top)
    flag = "BROKEN" if avg[top] > 0.5 and worst >= 0.5 else "no flag"
    print("  top deck {} share {:.3f}, worst matchup {:.2f}: {}".format(
        D[top], avg[top], worst, flag))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《非对称策略卡牌费用收益回归、牌库构筑超几何卡手率边界与优势策略循环克制》第 5 章「反事实遗憾最小化:从三牌暗押到卡组选择中的必胜坏牌组」(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-179 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026179,
  author    = {Fan, Yafu},
  title     = {Cost-Benefit Regression for Asymmetric Strategy Cards, Hypergeometric Brick-Rate Bounds in Deck Building and Cyclic Counters to Dominant Strategies},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论