5D Library › 演化博弈、直接与间接互惠及合作涌现的数学模型 · English

第四章 间接互惠:形象分与二阶声誉规范

帮助一个不会再见的人

直接互惠要求同一对个体反复相遇。人类社会中大量的帮助发生在陌生人之间:施助者不会从受助者那里得到回报,回报来自第三方,即「我帮了别人,别人看在眼里,所以有人帮我」。这就是间接互惠。

证据

形象分。 诺瓦克与西格蒙德 1998 年提出了一个间接互惠的模型:每个个体有一个公开的「形象分」,每帮助别人一次加分,每拒绝一次减分;「鉴别者」只帮助形象分不低于某个门槛的人。计算机模拟与解析模型都显示,只要知道对方形象分的概率 qq 超过代价与收益之比 c/bc/b,这样的合作就可以维持 (Nowak and Sigmund 1998)。韦德金德与米林斯基 2000 年在大学生中的实验支持了这一点:在无法直接回报的条件下,以前给予较多的人得到的帮助也较多 (Wedekind and Milinski 2000)。米林斯基、泽曼与克兰贝克 2002 年发现,把间接互惠的回合与公共品博弈交替进行,可以使公共品的捐献维持在高水平 (Milinski et al. 2002)。

形象分有一个漏洞。鉴别者拒绝帮助一个形象很差的背叛者,这本来是对背叛者的制裁;但在形象分的记账中,「拒绝」就是减分。于是正当的制裁者自己的形象也变差,其他鉴别者又会拒绝帮助他,他的形象再变差……制裁背叛者的人被当成了背叛者。莱马尔与哈默斯坦 2001 年在一个更现实的种群模型中发现,形象分策略并不稳定 (Leimar and Hammerstein 2001);潘查纳坦与博伊德 2003 年指出,问题出在形象分不区分「无故不帮」与「有理由不帮」,而经济学家萨格登 1986 年提出的「持守」(standing)概念恰好作出了这一区分 (Panchanathan and Boyd 2003; Sugden 1986)。

一阶与二阶评价

把声誉写成「好」与「坏」两种标签,评价规则可以按它看的信息分类。一阶规则只看施助者做了什么:帮助为好,拒绝为坏,这就是二元化的形象分。二阶规则还看受助者的标签。

推导

四种常见规则对施助者的评价(G 为好,B 为坏):

规则 帮助好人 拒绝好人 帮助坏人 拒绝坏人
形象分(一阶) G B G B
排斥(shunning) G B B B
简单持守(simple standing) G B G G
严格评判(stern judging) G B B G

四种规则的前两列相同;区别在于怎样评价与坏人打交道。简单持守下,拒绝坏人是正当的,帮助坏人也不算错;严格评判下,拒绝坏人是正当的,帮助坏人反而为坏;排斥下,与坏人打交道无论怎样都为坏。

证据

前八规范。 大槻与岩佐 2004 年系统地考察了一类更大的规则:评价时同时看施助者自己的标签、受助者的标签与施助者的行为(三阶评价),并配上相应的行动规则。在所有组合中,只有 8 种能使「鉴别者」的种群在有错误时维持高水平的合作并抵御其他策略的入侵,他们称之为「前八」(leading eight)(Ohtsuki and Iwasa 2004)。2006 年的论文分析了这八种规范的共同特点:好人帮助好人被评为好(维持合作);拒绝帮助好人被评为坏(识别背叛者);好人拒绝帮助坏人被评为好(正当的惩罚);坏人帮助好人可以被评为好(宽恕或道歉)。简单持守与严格评判都在前八之中,形象分与排斥不在 (Ohtsuki and Iwasa 2006)。

所以,框架中「拒绝帮助坏人不掉分」对前八规范都成立,这是它们与形象分的关键区别;「帮助坏人被判为坏」则只在严格评判等部分规范中成立,在简单持守下帮助坏人仍然为好。两者应当分开说。

程序:四种规则下的声誉动力学

下面的程序模拟一个 100 人的种群。每一步随机选一个施助者和一个受助者;帮助使施助者付出 c=1c = 1、受助者得到 b=3b = 3。鉴别者(DISC)帮助标签为好的人、拒绝标签为坏的人;永远合作者(ALLC)总是帮助;永远背叛者(ALLD)从不帮助。之后一个公共的观察者按规则给施助者贴上新标签,所有人都采用这个标签。有两种错误:想帮助却没有帮成(2%),以及观察者贴错标签(1%)。

推导

# Indirect reciprocity with public reputations (everyone shares one view, as in Ohtsuki
# & Iwasa 2004). N = 100 players; each step a random donor meets a random recipient and
# may give b = 3 at cost c = 1. DISC (discriminators) help Good recipients and refuse
# Bad ones; ALLC always help; ALLD never. An observer then labels the DONOR Good or Bad
# by a norm that looks at the action and the recipient's label (second-order norms;
# image scoring looks at the action only). Errors: an intended help fails with prob. e =
# 0.02; the observer assigns the wrong label with prob. mu = 0.01.
#                          help Good   refuse Good   help Bad   refuse Bad
#   image scoring              G            B            G           B
#   shunning                   G            B            B           B
#   simple standing            G            B            G           G
#   stern judging              G            B            B           G
import random
rng = random.Random(8)
KEYS = (("C", "G"), ("D", "G"), ("C", "B"), ("D", "B"))
NORMS = {name: dict(zip(KEYS, labels)) for name, labels in
         (("image scoring", "GBGB"), ("shunning", "GBBB"), ("simple standing", "GBGG"),
          ("stern judging", "GBBG"))}
B, C, E, MU = 3.0, 1.0, 0.02, 0.01

def run(norm, mix, steps=200000, N=100):
    types = [t for t, k in mix.items() for _ in range(k)]
    rep = ["G"] * N; pay = [0.0] * N
    good = {t: 0 for t in mix}; seen = {t: 0 for t in mix}
    zz_help = zz = 0
    for step in range(steps):
        d, r = rng.sample(range(N), 2)
        t = types[d]
        want = t == "ALLC" or (t == "DISC" and rep[r] == "G")
        act = "C" if want and rng.random() > E else "D"
        if act == "C": pay[d] -= C; pay[r] += B
        if t == "DISC" and types[r] == "DISC": zz += 1; zz_help += act == "C"
        lab = NORMS[norm][(act, rep[r])]
        if rng.random() < MU: lab = "B" if lab == "G" else "G"
        rep[d] = lab
        if step > steps // 4 and step % 50 == 0:
            for i in range(N): good[types[i]] += rep[i] == "G"; seen[types[i]] += 1
    av = {t: sum(p for p, ty in zip(pay, types) if ty == t) / mix[t] / steps * N
          for t in mix}
    return {t: good[t] / seen[t] for t in mix}, zz_help / zz, av

print("columns: share Good by strategy | DISC->DISC help rate | payoff per round")
for mix in ({"DISC": 90, "ALLD": 10}, {"DISC": 70, "ALLD": 30}, {"DISC": 80, "ALLC": 20},
            {"DISC": 70, "ALLD": 15, "ALLC": 15}):
    ks = list(mix)
    print("population " + ", ".join(f"{k} {v}" for k, v in mix.items()))
    for norm in NORMS:
        g, h, av = run(norm, mix)
        print(f"   {norm:16s} " + " ".join(f"{g[k]:5.2f}" for k in ks)
              + f"  |  {h:.2f}  |  " + " ".join(f"{av[k]:+.2f}" for k in ks))
# columns: share Good by strategy | DISC->DISC help rate | payoff per round
# population DISC 90, ALLD 10
#    image scoring     0.08  0.01  |  0.08  |  +0.14 +0.03
#    shunning          0.09  0.01  |  0.09  |  +0.16 +0.03
#    simple standing   0.97  0.11  |  0.95  |  +1.71 +0.28
#    stern judging     0.97  0.12  |  0.95  |  +1.70 +0.30
# population DISC 70, ALLD 30
#    image scoring     0.03  0.01  |  0.04  |  +0.05 +0.02
#    shunning          0.04  0.01  |  0.04  |  +0.05 +0.02
#    simple standing   0.98  0.24  |  0.96  |  +1.26 +0.51
#    stern judging     0.98  0.24  |  0.96  |  +1.26 +0.50
# population DISC 80, ALLC 20
#    image scoring     0.86  0.97  |  0.84  |  +1.73 +1.91
#    shunning          0.24  0.25  |  0.25  |  +0.94 +0.18
#    simple standing   0.97  0.97  |  0.95  |  +1.92 +1.90
#    stern judging     0.97  0.93  |  0.95  |  +1.93 +1.80
# population DISC 70, ALLD 15, ALLC 15
#    image scoring     0.45  0.01  0.97  |  0.44  |  +0.92 +0.47 +1.44
#    shunning          0.06  0.01  0.06  |  0.06  |  +0.52 +0.46 -0.43
#    simple standing   0.97  0.15  0.97  |  0.95  |  +1.61 +0.76 +1.44
#    stern judging     0.98  0.17  0.81  |  0.96  |  +1.64 +0.80 +1.11

第一组结果显示了形象分的漏洞:种群中有 90% 的鉴别者与 10% 的背叛者时,在形象分规则下鉴别者中只有 8% 被评为好,鉴别者之间的帮助率只有 8%,每轮平均收益 0.14,几乎与背叛者一样低。拒绝背叛者使鉴别者变坏,而拒绝变坏的鉴别者又使更多鉴别者变坏,合作在鉴别者之间瓦解了。排斥规则的结果几乎一样。在简单持守与严格评判下,鉴别者 97% 为好,彼此之间的帮助率为 95%,收益 1.71 与 1.70,远高于背叛者的 0.28 与 0.30。背叛者 30% 时,差别更明显。

程序中的形象分是最简单的二元版本,只记最近一次行为;诺瓦克与西格蒙德原文中的整数形象分更宽容一些。但漏洞的方向相同:只要拒绝本身被计为负分,正当的制裁就会受到惩罚。

第三组(鉴别者 80%,永远合作者 20%,没有背叛者)显示了另一面。形象分下,永远合作者的收益(1.91)高于鉴别者(1.73),因为鉴别者偶尔会因错误拒绝别人而被扣分。简单持守下两者几乎相同(1.92 与 1.90):在没有背叛者的时候,「无条件的好人」与「有条件的好人」无从区分。严格评判下永远合作者被评为好的比例降到 93%、收益降到 1.80,因为他们帮助了因错误而变坏的人。这说明了严格评判的作用:它惩罚对坏人的无条件帮助,因此能阻止永远合作者在种群中慢慢增加,而永远合作者的增加正是背叛者入侵的前奏。

推导

# The model of p08 in an infinite population: the share of Good players of each strategy
# solves a fixed-point equation, payoffs follow, and the replicator equation moves the
# strategy shares (DISC, ALLC, ALLD). Same norms and errors (b = 3, c = 1, e = 0.02, mu
# = 0.01), plus rare mutation: each strategy share drifts towards 1/3 at rate 0.002, so
# a strategy that has died out can come back. A start "ends cooperative" if the final
# state has cooperation above 80 % (share of interactions in which help is given).
NORMS = {"image scoring":   {("C", "G"): 1, ("D", "G"): 0, ("C", "B"): 1, ("D", "B"): 0},
         "shunning":        {("C", "G"): 1, ("D", "G"): 0, ("C", "B"): 0, ("D", "B"): 0},
         "simple standing": {("C", "G"): 1, ("D", "G"): 0, ("C", "B"): 1, ("D", "B"): 1},
         "stern judging":   {("C", "G"): 1, ("D", "G"): 0, ("C", "B"): 0, ("D", "B"): 1}}
B, C, E, MU = 3.0, 1.0, 0.02, 0.01

def reputations(x, nm, g0=None):
    gd, gc, ga = g0 or (1.0, 1.0, 1.0)
    for _ in range(500):
        g = x[0] * gd + x[1] * gc + x[2] * ga             # chance a recipient is Good
        def good(pc_g, pc_b):                              # pc_* = P(help) vs Good / Bad
            pg = g * (pc_g * nm[("C", "G")] + (1 - pc_g) * nm[("D", "G")]) \
                + (1 - g) * (pc_b * nm[("C", "B")] + (1 - pc_b) * nm[("D", "B")])
            return (1 - MU) * pg + MU * (1 - pg)
        new = (good(1 - E, 0.0), good(1 - E, 1 - E), good(0.0, 0.0))
        if max(abs(a - b) for a, b in zip(new, (gd, gc, ga))) < 1e-12: break
        gd, gc, ga = new
    return gd, gc, ga

def payoffs(x, nm, g0=None):
    gd, gc, ga = reputations(x, nm, g0)
    g = x[0] * gd + x[1] * gc + x[2] * ga
    recv = lambda gi: B * (1 - E) * (x[1] + x[0] * gi)
    return (recv(gd) - C * (1 - E) * g, recv(gc) - C * (1 - E), recv(ga)), (gd, gc, ga)

def evolve(x, nm, t=1500.0, dt=0.2, m=0.002):
    reps = None
    for _ in range(int(t / dt)):
        p, reps = payoffs(x, nm, reps)
        av = sum(a * b for a, b in zip(x, p))
        y = [max(0.0, xi + dt * (xi * (pi - av) + m * (1 / 3 - xi)))
             for xi, pi in zip(x, p)]
        s = sum(y); y = [v / s for v in y]
        done = max(abs(a - b) for a, b in zip(x, y)) < 1e-9
        x = y
        if done: break
    coop = (1 - E) * (x[0] * (x[0] * reps[0] + x[1] * reps[1] + x[2] * reps[2]) + x[1])
    return x, coop

print("check against p08 (DISC 90 %, ALLD 10 %): share Good of DISC, ALLD")
for name, nm in NORMS.items():
    gd, gc, ga = reputations([0.9, 0.0, 0.1], nm)
    print(f"   {name:16s} {gd:.2f}  {ga:.2f}")
print("share of 36 starting mixes that end cooperative;\n   end state from DISC 0.6,"
      " ALLC 0.2, ALLD 0.2:")
starts = [[i / 10, j / 10, 1 - (i + j) / 10]
          for i in range(1, 10) for j in range(1, 10 - i)]
for name, nm in NORMS.items():
    ok = 0
    for s0 in starts:
        x, coop = evolve(s0, nm)
        ok += coop > 0.8
    x, coop = evolve([0.6, 0.2, 0.2], nm)
    print(f"   {name:16s} {ok / len(starts):.2f}   DISC {x[0]:.2f} ALLC {x[1]:.2f}"
          f" ALLD {x[2]:.2f}, cooperation {coop:.2f}")
# check against p08 (DISC 90 %, ALLD 10 %): share Good of DISC, ALLD
#    image scoring    0.08  0.01
#    shunning         0.08  0.01
#    simple standing  0.97  0.12
#    stern judging    0.97  0.12
# share of 36 starting mixes that end cooperative;
#    end state from DISC 0.6, ALLC 0.2, ALLD 0.2:
#    image scoring    0.00   DISC 0.11 ALLC 0.00 ALLD 0.89, cooperation 0.00
#    shunning         0.00   DISC 1.00 ALLC 0.00 ALLD 0.00, cooperation 0.24
#    simple standing  0.53   DISC 0.98 ALLC 0.02 ALLD 0.00, cooperation 0.95
#    stern judging    0.61   DISC 0.99 ALLC 0.01 ALLD 0.00, cooperation 0.95

第二段程序在无限种群中求出同一个模型的声誉平衡,并让三种策略的比例按复制子方程演化(加上很少的突变)。它与第一段程序吻合:90% 鉴别者、10% 背叛者时,形象分下鉴别者 8% 为好,简单持守与严格评判下 97%。在 36 个初始比例中,形象分与排斥下没有一个最终维持合作;简单持守下 53%、严格评判下 61% 的初始比例走向以鉴别者为主、合作率 95% 的状态。从「鉴别者 0.6、永远合作者 0.2、背叛者 0.2」出发,形象分下种群最终 89% 为背叛者;排斥下鉴别者占满种群,但因为错误不断制造「坏人」而合作率只有 24%;简单持守与严格评判下合作率为 95%。

四种评价规则下鉴别者(上顶点)、永远合作者(左下)与永远背叛者(右下)比例的演化轨迹。小点为起点,大点为终点。形象分下轨迹先滑向永远合作者一侧,然后被背叛者占领;简单持守与严格评判下,起点中鉴别者足够多时种群走向鉴别者顶点。

注

公开评价的假设。 前八规范的结论建立在一个强假设上:所有人对每个人的好坏看法一致,评价是公开的、共享的。希尔贝、施密德、特卡德莱茨、查特吉与诺瓦克 2018 年发现,当每个人私下观察、独立评价,并且观察不完整或有误差时,个体之间的看法很快出现分歧,前八规范中的大多数不再能维持合作 (Hilbe et al. 2018)。所以,「采用二阶评价规则」本身不足以保证合作;评价信息怎样被收集与共享同样重要。第八章的核查表把这一点列为单独的一项。

「以直报怨」:类比而非等同

间接互惠的规范与古典伦理之间有一个容易引起联想的对照。《论语·宪问》中,有人问:「以德报怨,何如?」孔子回答:「何以报德?以直报怨,以德报德。」(第七章录全文。)「以德报德」与「好人帮助好人为好」相似;「以直报怨」与「拒绝坏人为好」也有几分相似:对于伤害自己的人,不以恩德回报,但也不以怨恨回报,而是以「直」相待。

这种对照只是结构上的类比。孔子讨论的是君子个人对待恩怨的态度,「直」在朱熹的解释中是「爱憎取舍,一以至公而无私」,是一种公正的心态,不是一条「拒绝帮助」的行动规则;而声誉规范讨论的是一个群体中由第三方执行的评价制度,它的效果是用种群中策略频率的变化来衡量的。把孔子说成「早已提出了二阶声誉规范」,既误解了孔子,也误解了模型。

假说 · 尚未验证

正当拒绝标注假说(尚未确立)。作者推测,在一个依赖成员互助的线上社群中,如果公开的贡献记录把「拒绝向被标记为违规者提供帮助」与「无故拒绝」区分开来,成员对违规者的制裁意愿会提高,而对守规成员的帮助不会减少。可以检验的预测是:在随机分配到两种记录方式的社群中,区分正当拒绝的社群里,违规者在被标记后得到的帮助下降更多,而守规成员之间的互助率不低于对照社群。如果两种社群对违规者的帮助没有差别,或者区分正当拒绝使守规成员之间的互助减少,这一假说就被推翻。第八章的试验检验与之相关的一个问题。那项试验的标签只用一个口径,即上表中的简单持守,并取最严格的正当化:对坏声誉者不捐献为好,对好声誉者不捐献为坏,向坏人捐献不扣分;在 3 人小组中,只有同组另外两人都为坏时不捐献才算正当。它不是严格评判(帮助坏人不扣分),也不是形象分(形象分不区分正当与不正当的不捐献)。

本章小结

间接互惠让合作可以发生在不会再见的陌生人之间,条件是声誉足够为人所知(q>c/bq > c/b)。一阶的形象分有一个漏洞:拒绝帮助坏人也被扣分,于是正当的制裁者被当成背叛者,合作在有背叛者时瓦解。二阶评价把「无故拒绝」与「正当拒绝」区分开来;大槻与岩佐的前八规范都具有这一性质,其中严格评判还把帮助坏人评为坏,从而阻止无条件合作者的增加。这些结论依赖于评价公开、共享的假设。下一章讨论另一种不需要声誉的机制:谁与谁相邻。

参考文献

Hilbe, Christian, Laura Schmid, Josef Tkadlec, Krishnendu Chatterjee, and Martin A. Nowak. 2018. “Indirect Reciprocity with Private, Noisy, and Incomplete Information.” Proceedings of the National Academy of Sciences 115 (48): 12241–46. https://doi.org/10.1073/pnas.1810565115.
Leimar, Olof, and Peter Hammerstein. 2001. “Evolution of Cooperation Through Indirect Reciprocity.” Proceedings of the Royal Society B: Biological Sciences 268 (1468): 745–53. https://doi.org/10.1098/rspb.2000.1573.
Milinski, Manfred, Dirk Semmann, and Hans-Jürgen Krambeck. 2002. “Reputation Helps Solve the ‘Tragedy of the Commons’.” Nature 415 (6870): 424–26. https://doi.org/10.1038/415424a.
Nowak, Martin A., and Karl Sigmund. 1998. “Evolution of Indirect Reciprocity by Image Scoring.” Nature 393 (6685): 573–77. https://doi.org/10.1038/31225.
Ohtsuki, Hisashi, and Yoh Iwasa. 2004. “How Should We Define Goodness? Reputation Dynamics in Indirect Reciprocity.” Journal of Theoretical Biology 231 (1): 107–20. https://doi.org/10.1016/j.jtbi.2004.06.005.
Ohtsuki, Hisashi, and Yoh Iwasa. 2006. “The Leading Eight: Social Norms That Can Maintain Cooperation by Indirect Reciprocity.” Journal of Theoretical Biology 239 (4): 435–44. https://doi.org/10.1016/j.jtbi.2005.08.008.
Panchanathan, Karthik, and Robert Boyd. 2003. “A Tale of Two Defectors: The Importance of Standing for Evolution of Indirect Reciprocity.” Journal of Theoretical Biology 224 (1): 115–26. https://doi.org/10.1016/S0022-5193(03)00154-1.
Sugden, Robert. 1986. The Economics of Rights, Co-Operation and Welfare. Blackwell.
Wedekind, Claus, and Manfred Milinski. 2000. “Cooperation Through Image Scoring in Humans.” Science 288 (5467): 850–52. https://doi.org/10.1126/science.288.5467.850.

Fan, Y. (2026). Evolutionary Game Theory, Direct and Indirect Reciprocity and the Emergence of Cooperation: From Hamilton's Rule and Axelrod's Tournaments to Reputation and Costly Punishment, Read Beside Classical Confucian and Mohist Ethics [演化博弈、直接与间接互惠及合作涌现的数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23788551 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).