5D Library › Books › 5DL-B-055

四人不完全信息随机博弈、贝叶斯读牌与反事实遗憾最小化

从牌池马尔可夫状态转移、防守弃胡的最优停时到多人博弈的资金凯利分配及古典兵家「虚实奇正」考辨

Imperfect-Information Stochastic Games, Bayesian Hand-Reading and Counterfactual Regret Minimisation in Four-Player Tile Dynamics: From Markov Transitions of the Tile Pool and the Optimal Stopping of Defensive Folding to Kelly Allocation in a Multiplayer Game, with a Textual Examination of the Orthodox and the Unorthodox in the Classical Military Writers

  • 范雅福 Fan Yafu
  • Version 1.0.0 · 2026-11-07
  • CC-BY-NC-ND-4.0
  • 5DL-B-055
DOI10.5281/zenodo.24020464

摘要

一张四人牌桌是博弈论里最不讨人喜欢的那一类对象:不完全信息、随机、四方、而且在金钱上是负和。本书把这类博弈写成数学,并逐段检验流行的策略推理。它不是赌博指南:第八章算出,一个有合理优势的玩家一年打 3,120 手、抽水 0.010 时期望收益为 +8.6 个买入,却仍有 36% 的人年终亏损;而没有优势时任何抽水都使期望为负,无论用什么资金管理规则。

第一章给出规模:13 张手牌的不同多重集约有 9.85 × 10¹⁰ 个,另外三家的配置约 5.0 × 10⁴⁶ 种,所以枚举信息集是范畴错误。第二章把牌池写成超几何推断:两面听相对单骑的优势只在**后期**接近 2(剩 2 次摸牌时 1.94),**前期**只有 1.38。第三章把筋牌推导成贝叶斯演绎:它使某张牌安全 3.6 倍,而非完全安全。第四章用动态规划算向聴,并指出单人「赛跑」模型把胜率高估 1.36 到 1.79 倍,误差随距离增大。第五章把弃胡写成最优停时:对手自摸的项**相消**,所以弃牌的理由只落在放铳概率上。

第六章在两人零和博弈上验证反事实遗憾最小化(可利用度从 0.667 降到 0.0013,博弈值 −0.055554,理论值 −1/18),并在三人博弈上**诚实地报告两个否定结果**:可利用度照样降到 4 × 10⁻⁶,四种简单偏离也没能把玩家压到均衡值以下。凯利增长率曲线在最优点附近近乎**对称**(半凯利保 75.3%、1.5 倍凯利保 75.7%),所以降低注码的理由是估计误差而不是曲线形状。第八章给出四步核查表与一项预注册的重复对局策略比较,只在首次分歧的决策上计分:最小关注效应 0.010 个买入/计分决策,需 12,000 个计分决策,判定规则写成代码;结论只对所用的固定对手池成立。书中以标注框区分已发表的证据、可运行的 Python 计算(16 段程序)、原典与作者的模型和假说;除注明的文献数据外,书中数据都是合成的。

关键词:不完全信息博弈;随机博弈;扩展式博弈;信息集;超几何推断;贝叶斯推断;读牌;向聴数;动态规划;吸收马尔可夫链;竞争风险;最优停时;马尔可夫决策过程;反事实遗憾最小化;遗憾匹配;纳什均衡;多人博弈;可利用度;混合策略;凯利准则;遍历性;破产概率;重复对局设计;方差缩减;预注册;赌博危害

Abstract

A four-person table is the least agreeable class of object in game theory: imperfect information, stochastic, four-sided and negative-sum in money. This book writes that class of game as mathematics and tests popular strategic reasoning segment by segment. It is not a gambling guide: Chapter 8 computes that a player with a plausible edge, playing 3,120 hands a year at a fee of 0.010, expects +8.6 buy-ins while 36 % of such players still finish down, and that with no edge any fee makes the expectation negative whatever the staking rule.

Chapter 1 gives the scale: about 9.85 × 10¹⁰ distinct 13-tile hands and about 5.0 × 10⁴⁶ configurations of the other three hands, so enumerating information sets is a category error. Chapter 2 treats the tile pool as hypergeometric inference: a two-sided wait's advantage over a one-sided wait approaches 2 only **late** in the hand (1.94 with two draws left) and is 1.38 **early**. Chapter 3 derives suji as a Bayesian deduction that makes a tile 3.6 times safer, not safe. Chapter 4 computes shanten by dynamic programming and shows that a one-player race model overstates the win probability by factors of 1.36 to 1.79, the error growing with distance. Chapter 5 writes folding as optimal stopping: the opponent's self-draw term **cancels**, so the case for folding rests on the deal-in probability alone.

Chapter 6 verifies counterfactual regret minimisation on a two-player zero-sum game (exploitability from 0.667 to 0.0013, value −0.055554 against −1/18) and **honestly reports two negative results** on a three-player game: exploitability still falls to 4 × 10⁻⁶, and four simple deviations failed to push the player below the equilibrium value. The Kelly growth curve is nearly **symmetric** (half-Kelly keeps 75.3 %, 1.5× Kelly 75.7 %), so the reason to stake less is estimation error, not the curve's shape. Chapter 8 gives a four-step checklist and a pre-registered duplicate-design policy comparison scored at the first differing decision: smallest effect of interest 0.010 buy-ins per scored decision, 12,000 scored decisions, the decision rule written as code; the conclusion holds only against the fixed opponent pool used. Marked boxes separate published evidence, runnable Python calculations (16 programs), classical sources and the author's models and hypotheses; apart from cited literature data, all data in the book are synthetic.

Keywords: imperfect information games; stochastic games; extensive-form games; information sets; hypergeometric inference; Bayesian inference; hand reading; shanten; dynamic programming; absorbing Markov chain; competing risk; optimal stopping; Markov decision process; counterfactual regret minimisation; regret matching; Nash equilibrium; multiplayer games; exploitability; mixed strategies; Kelly criterion; ergodicity; risk of ruin; duplicate design; variance reduction; pre-registration; gambling harm

目录 · Contents

  • 序:最不讨人喜欢的那一类博弈English 中文Preface: The Least Agreeable Class of Game
  • 第一章 规模:为什么不能把信息集列出来English 中文Chapter 1. Scale: Why the Information Sets Cannot Be Listed
  • 第二章 牌池:超几何推断与一个方向相反的结果English 中文Chapter 2. The Tile Pool: Hypergeometric Inference and a Result That Runs Backwards
  • 第三章 读牌:筋是一条演绎,不是一条规律English 中文Chapter 3. Hand Reading: Suji Is a Deduction, Not a Regularity
  • 第四章 向聴、进张与赛跑:忽略三家会高估胜率English 中文Chapter 4. Shanten, Acceptance and the Race: Ignoring Three Players Overstates the Win
  • 第五章 弃胡:最优停时,与一个被说反的结论English 中文Chapter 5. Folding: Optimal Stopping, and a Conclusion Stated Backwards
  • 第六章 博弈的解与钱English 中文Chapter 6. Solving the Game, and the Money
  • 第七章 原典并读:虚实、奇正与战争的迷雾English 中文Chapter 7. Classics Read Side by Side: The Substantial and the Empty, the Orthodox and the Unorthodox, and the Fog of War
  • 第八章 实践与预注册检验English 中文Chapter 8. Practice and a Pre-Registered Test

How to cite · 引用

Fan, Y. (2026). Imperfect-Information Stochastic Games, Bayesian Hand-Reading and Counterfactual Regret Minimisation in Four-Player Tile Dynamics: From Markov Transitions of the Tile Pool and the Optimal Stopping of Defensive Folding to Kelly Allocation in a Multiplayer Game, with a Textual Examination of the Orthodox and the Unorthodox in the Classical Military Writers [四人不完全信息随机博弈、贝叶斯读牌与反事实遗憾最小化] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24020464

BibTeX
@book{fan2026imperfect,
  title = {{Imperfect-Information Stochastic Games, Bayesian Hand-Reading and Counterfactual Regret Minimisation in Four-Player Tile Dynamics}: From Markov Transitions of the Tile Pool and the Optimal Stopping of Defensive Folding to Kelly Allocation in a Multiplayer Game, with a Textual Examination of the Orthodox and the Unorthodox in the Classical Military Writers},
  author = {Fan, Yafu},
  year = {2026},
  publisher = {5D Library},
  edition = {Version 1.0.0},
  doi = {10.5281/zenodo.24020464},
  url = {https://doi.org/10.5281/zenodo.24020464},
  note = {Chinese title: 四人不完全信息随机博弈、贝叶斯读牌与反事实遗憾最小化}
}
DOIhttps://doi.org/10.5281/zenodo.24020464
DOI(全部版本 all versions)https://doi.org/10.5281/zenodo.24020463
ORCIDhttps://orcid.org/0009-0004-3519-1203
5D Library ID5DL-B-055
版本 Version1.0.0 · 中英双语版 Bilingual edition
出版日期 Date2026-11-07
网址 Onlinehttps://5dlibrary.com/books/imperfect-information-tile-games/
SHA-256 · imperfect-information-tile-games_bilingual_v1.0.0.pdfb1566e5a6c43f7a8e271a167fbbb7288420a3fa780abfe9d80a0b48c2f9d436d
SHA-256 · imperfect-information-tile-games_en_v1.0.0.pdfc4e2651264185db626957798922bfaf7c0aba4ae9a1f82c4956f43bc590ce1b1
SHA-256 · imperfect-information-tile-games_zh_v1.0.0.pdfd23b51d323b5a22b129a749be65faec67e12768d9a990802b9bc7de324759e4e
Licence · 许可Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0). https://creativecommons.org/licenses/by-nc-nd/4.0/

Disclosure. The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base, and therefore has a commercial interest in the subject of this book. The book promotes no course, card deck, software or service, provides no link to or recommendation of any gambling platform, agent or funding service, and accepts no sponsorship from any such organisation. This book is not a gambling guide: it does not advise anyone to stake money on the games it discusses, the last section of Chapter 8 computes why staking on an unmeasured edge is a quantifiable error, and it states that gambling can cause serious financial and personal harm and can become compulsive. The Chapter 8 trial is a comparison of policies between software agents, to be run and analysed by researchers independent of the author's business, involves no real-money wagering, and does not recruit from the author's community, students or customers; all the code and the study protocol are published openly for anyone to check.