5D Library › 算法博弈论、布雷斯悖论与流量分配的无秩序代价 · English

第八章 实践与预注册试验

注

本章的核查表是一个思考工具,不是管理规范、营销方法或投资建议。 示例是一个虚构的「青禾工作室」,所有人名、品牌、渠道与数字都是编造的,只用来演示计算方法,不来自任何真实的组织、平台或客户。任何真实团队的流程设计、预算分配与合作安排,都需要依据自己的数据并在必要时咨询相关专业人员。

第一部分 四步去拥堵与分流核查表

前六章的模型可以整理成四个问题。它们帮助一个团队找出自己的流程、预算与合作中,哪些地方的「每个人都做对自己最好的事」可能让整体变差。

  1. 单点瓶颈的布雷斯捷径审计。 流程中有没有一条「快速通道」,把原本分别承担不同负荷的两个关键人连成了所有事务的必经之路(第一章)?关掉它,平均处理时间会不会反而缩短?
  2. 红海赛道的边际拥堵税计算。 同一团队的几个项目是否在争夺同一批受众、同一组关键词?分配预算时看的是每个渠道的平均获客成本,还是多投一单位时的边际成本(第三、四章)?两者之差就是自己给自己征的拥堵税。
  3. 在制品上限节流。 同时打开的任务有没有上限?任务之间切换的代价越大,不设上限越危险(第七章《度地》的「满则后推前」只是类比)。
  4. 合作撮合的稳定匹配检查。 现有的合作配对中有没有阻挡对:一个创作者与一个品牌,彼此都更愿意与对方合作(第六章)?如果用延迟接受的方式重新配对,由哪一方提议?

推导

把核查表用在虚构的「青禾工作室」上。第 1 步:两位资深审核人 W 与 L 各需 4x4x 天(xx 为经手请求的份额),标准质检与标准合规各需 3 天;原本请求走「W → 质检」或「合规 → L」,「快速通道」允许 W 审完直接交给 L。第 2 步:团队每周要获得 3,000 位顾客,三个渠道的单位获客成本随团队自己的投放量上升:热门关键词 6+10q6 + 10q,中等渠道 9+3q9 + 3q,小众渠道 12+q12 + q(qq 以千人计)。比较「使各渠道平均成本相等」与「使边际成本相等」两种分配。第 3 步:任务每周到达 8 件,一个人专心做一件事每周能完成 10 件,但每多开一件任务,切换使效率降低 3%;比较不同的在制品上限 KK(超出上限的任务在看得见的待办队列中等待)。第 4 步:4 位创作者与 4 个品牌的偏好与当前配对,检查阻挡对,并比较双方各自提议的延迟接受结果。

推导

# The four-step checklist applied to a FICTIONAL team ("Studio Qinghe"); every name and
# number is invented to show the arithmetic, not taken from any real organisation or
# client.

# Step 1. Braess audit of an approval "fast track". Requests flow from intake to
# release. Senior reviewers W and L each take 4x days (x = share of weekly requests they
# handle); standard QA and standard compliance take 3 days each. Route A: W -> QA; route
# B: compliance -> L. A fast-track rule lets work go W -> L directly, skipping both
# standard checks. Equilibrium = minimum of the potential 2w^2 + 2l^2 + 3a + 3b (w, l =
# loads on W and L).
for ft in (False, True):
    grid = [(i / 200, j / 200) for i in range(201) for j in range(201 - i)
            if ft or i + j == 200]
    phi = lambda g: 2 * (1 - g[1]) ** 2 + 2 * (1 - g[0]) ** 2 + 3 * g[0] + 3 * g[1]
    a, b = min(grid, key=phi)
    z, w, l = 1 - a - b, 1 - b, 1 - a
    t = a * (4 * w + 3) + b * (3 + 4 * l) + z * (4 * w + 4 * l)
    print(f"step 1: fast track {'on ' if ft else 'off'}: average days {t:.2f};"
          f" shares A {a:.2f}, B {b:.2f}, fast track {z:.2f}")

# Step 2. The team's own campaigns crowd each other. Cost per acquisition in channel k
# rises with the team's weekly acquisitions q_k (thousands): CPA_k = a_k + b_k q_k.
# Target: 3,000.
CH = {"hot keyword": (6.0, 10.0), "mid channel": (9.0, 3.0), "niche": (12.0, 1.0)}
def allocate(marginal, Q=3.0):
    lo, hi = 0.0, 100.0                       # common level of average or marginal CPA
    for _ in range(100):
        m = (lo + hi) / 2
        q = {k: max(0.0, (m - a) / (b * (2 if marginal else 1)))
             for k, (a, b) in CH.items()}
        if sum(q.values()) > Q: hi = m
        else: lo = m
    return q
for marginal in (False, True):
    q = allocate(marginal)
    spend = sum(q[k] * (a + b * q[k]) for k, (a, b) in CH.items())
    rule = "equal marginal CPA" if marginal else "equal average CPA "
    print(f"step 2: {rule}: " + ", ".join(f"{k} {v:.2f}" for k, v in q.items())
          + f"; spend {spend:.2f}")
q = allocate(False)
print("        congestion tax q*b, average rule: "
      + ", ".join(f"{k} {q[k] * b:.2f}" for k, (a, b) in CH.items()))

# Step 3. A cap on work in progress (WIP). Tasks arrive at 8 a week; one person
# finishing one task at a time does 10 a week, but each extra open task costs 3 % of
# capacity in switching (rate 10 / (1 + 0.03 (open - 1))). Tasks beyond the cap K wait
# in a visible backlog.
lam = 8.0
mu = lambda n: 10 / (1 + 0.03 * (n - 1))
print("step 3:  WIP cap K   open tasks  in backlog     days request->done"
      "  days start->done")
for K in (1, 2, 3, 5, 8, None):
    p, stable = [1.0], True
    for n in range(1, 5000):
        p.append(p[-1] * lam / mu(n if K is None else min(n, K)))
        if p[-1] > 1e12: stable = False; break
    if not stable:
        print(f"        {'none':>8}    no steady state: open tasks pile up without limit")
        continue
    s = sum(p); p = [x / s for x in p]
    L = sum(n * x for n, x in enumerate(p))
    O = sum(min(n, K) * x for n, x in enumerate(p))
    print(f"        {K:>8}    {O:5.2f}       {L - O:5.2f}          {7 * L / lam:5.1f}"
          f"               {7 * O / lam:5.1f}")

# Step 4. Stable-matching check of the current creator-brand pairings.
CRE = ["Ana", "Bo", "Chen", "Dee"]; BR = ["tea", "gear", "skin", "books"]
PC = {"Ana": ["tea", "gear", "skin", "books"], "Bo": ["gear", "skin", "tea", "books"],
      "Chen": ["skin", "tea", "gear", "books"], "Dee": ["books", "tea", "gear", "skin"]}
PB = {"tea": ["Bo", "Chen", "Ana", "Dee"], "gear": ["Chen", "Ana", "Bo", "Dee"],
      "skin": ["Ana", "Bo", "Chen", "Dee"], "books": ["Dee", "Ana", "Bo", "Chen"]}
now = {"Ana": "books", "Bo": "tea", "Chen": "gear", "Dee": "skin"}
def blocking(m):
    inv = {b: c for c, b in m.items()}
    return [(c, b) for c in CRE for b in BR if b != m[c]
            and PC[c].index(b) < PC[c].index(m[c])
            and PB[b].index(c) < PB[b].index(inv[b])]
def da(prop, pp, rp):                            # deferred acceptance, "prop" proposes
    held, nxt, free = {}, {x: 0 for x in prop}, list(prop)
    while free:
        x = free.pop(0); y = pp[x][nxt[x]]; nxt[x] += 1
        if y not in held: held[y] = x
        elif rp[y].index(x) < rp[y].index(held[y]): free.append(held[y]); held[y] = x
        else: free.append(x)
    return held
pairs = lambda m: ", ".join(f"{c}-{m[c]}" for c in CRE)
bp = lambda m: ", ".join(f"{c}-{b}" for c, b in blocking(m)) or "none"
print(f"step 4: current pairs {pairs(now)}\n        blocking pairs: {bp(now)}")
for lab, m in (("creators propose", {c: b for b, c in da(CRE, PC, PB).items()}),
               ("brands propose  ", da(BR, PB, PC))):
    m = {c: m[c] for c in CRE}; inv = {b: c for c, b in m.items()}
    print(f"        {lab}: {pairs(m)}; blocking pairs: {bp(m)}"
          "\n           ranks of partners: "
          f"creators {[PC[c].index(m[c]) + 1 for c in CRE]}, "
          f"brands {[PB[b].index(inv[b]) + 1 for b in BR]}")
# step 1: fast track off: average days 5.00; shares A 0.50, B 0.50, fast track 0.00
# step 1: fast track on : average days 6.00; shares A 0.25, B 0.25, fast track 0.50
# step 2: equal average CPA : hot keyword 0.70, mid channel 1.33, niche 0.98; spend 38.93
# step 2: equal marginal CPA: hot keyword 0.45, mid channel 1.01, niche 1.53; spend 37.73
#         congestion tax q*b, average rule: hot keyword 6.98, mid channel 3.98, niche 0.98
# step 3:  WIP cap K   open tasks  in backlog     days request->done  days start->done
#                1     0.80        3.20            3.5                 0.7
#                2     1.50        3.16            4.1                 1.3
#                3     2.14        3.34            4.8                 1.9
#                5     3.47        4.70            7.1                 3.0
#                8     6.47       21.43           24.4                 5.7
#             none    no steady state: open tasks pile up without limit
# step 4: current pairs Ana-books, Bo-tea, Chen-gear, Dee-skin
#         blocking pairs: Ana-skin, Bo-skin, Chen-skin, Dee-books
#         creators propose: Ana-tea, Bo-gear, Chen-skin, Dee-books; blocking pairs: none
#            ranks of partners: creators [1, 1, 1, 1], brands [3, 3, 3, 1]
#         brands propose  : Ana-skin, Bo-tea, Chen-gear, Dee-books; blocking pairs: none
#            ranks of partners: creators [3, 3, 3, 1], brands [1, 1, 1, 1]

第 1 步:没有快速通道时,请求各走一半,平均 5.00 天;开放快速通道后,一半请求走 W → L,W 与 L 各自承担 75% 的请求,三条路线都是 6.00 天,所有请求慢了 20%。这是第一章布雷斯网络的翻版:快速通道对每个请求都「更快」,直到所有人都用它。结论是检查快速通道是否应只对少数真正紧急的请求开放,而不是取消审核。

第 2 步:按平均成本相等分配时,热门关键词带来 700 位顾客,总花费 38.93(以千计的货币单位,下同);按边际成本相等分配时,热门关键词降到 450 位、小众渠道升到 1,530 位,总花费 37.73,少约 3%。在平均成本规则下,热门关键词上的「拥堵税」q×bq \times b 为 6.98,是小众渠道(0.98)的 7 倍:多往热门关键词投一单位,不仅自己付出平均成本,还推高了同一团队在这一渠道上所有其他投放的成本。节省的幅度不大,与第二章「线性成本下损失有限」的结论一致;但方向是确定的。

第 3 步:上限为 1 时,从提出请求到完成平均 3.5 天,开始做之后只需 0.7 天;上限为 3 时为 4.8 天;上限为 8 时升到 24.4 天,其中只有 5.7 天在做,其余都在待办队列里等待:切换使完成速度只比到达速度略快;不设上限时,完成速度随打开的任务增多而下降,最终低于到达速度,系统没有稳定状态,积压无限增长。这个模型中上限越小越好,是因为它没有「等待别人的反馈」这类需要并行的理由;真实工作中合适的上限要用自己的数据确定。切换成本的大小(3%)是示意值。排队的平均数量与等待时间之间的关系服从利特尔法则 (Little 1961)。

第 4 步:当前的配对有四个阻挡对,例如 Ana 与 skin 彼此都更喜欢对方。创作者提议时,四位创作者都得到第一志愿,三个品牌得到第三志愿;品牌提议时正好相反。两种结果都没有阻挡对,都是稳定的;选哪一个,是对谁更有利的规则选择,应当事先说明。所有名字与偏好都是编造的。

第二部分 预注册随机对照试验方案

问题

在一个含有会触发布雷斯悖论的捷径的多轮路线选择博弈中,每轮结束时除了各路线的通行时间之外,再向参与者显示「你这一轮给其他人增加了多少延迟」,并按这一数值扣减拥堵积分,能否减少小组陷入布雷斯低效均衡的比例?

证据

已有的实验。 拉波波特、库格勒、杜加尔与吉谢斯 2009 年让较大的小组在有、无附加路段的两种网络中反复选择路线,结果否定了「布雷斯悖论只是边缘现象、会随经验消失」的看法:有经济激励的参与者随经验收敛到扩展网络中的均衡路线,尽管收益因此大幅下降 (Rapoport et al. 2009)。摩根、奥尔森与塞夫顿 2009 年发现,在按布雷斯方式改变的网络中,参与者调整路线后收益下降,总体流量接近均衡水平,但有系统性的偏离,可以用简单的强化学习模型解释 (Morgan et al. 2009)。泽尔滕等人 2007 年让每组 18 人在两条路之间反复选择,平均人数接近均衡,但波动一直持续;只知道自己通行时间的一组换路次数显著更多 (Selten et al. 2007)。吉谢斯与拉波波特 2012 年比较了「只知道自己收益」与「知道所有人的路线与收益」两种信息条件,两种条件下总体路线选择都收敛到均衡 (Gisches and Rapoport 2012)。

这些结果说明,单纯提供更多关于路况的信息不足以让参与者离开布雷斯均衡。本试验检验的是另一种干预:让每个人看到并承担自己的外部性。第三章的程序显示,按外部性扣分后,「全走桥」不再是均衡。

设计

要素 规定
设计 两组平行、以 6 人小组为单位随机化的对照试验,按实施地点分层,1:1 分配
参与者 年满 18 岁的成年人,通过大学与研究机构的公开招募参加,按实验经济学惯例获得与任务收益挂钩的报酬;不从作者的社群、学员或客户(包括作者企业的客户)中招募
排除 曾参加过路线选择或拥堵博弈实验者;曾参加本研究预试验者
任务 40 轮。每轮每人从上路 U、下路 L 与桥路 Z 中选一条(第三章的网络:拥堵路段每人 10,固定路段 70,桥 0);每轮收益 = 200 − 通行时间(A 组再减去拥堵积分)。规则在开始前以书面与示例说明,两组相同
两组共同的界面 每轮结束后显示:各路线本轮的通行时间、自己的通行时间与收益、累计收益
A 组(外部性显示 + 积分) 另显示「你本轮给其他人增加的延迟」,即你走过的每段拥堵路上其他使用者的人数 × 10,并从收益中扣减同样数目的拥堵积分;附两分钟的中性说明(「拥堵路段上每多一个人,其他每人慢 10」),不提示应走哪条路
B 组(对照) 只有共同界面
盲法 参与者知道自己的界面,但不知道研究假设与另一组的界面;主要结局由程序自动计算;分析人员在分析计划锁定前不知道组别
主要结局 陷入布雷斯均衡:第 31—40 轮中至少有 8 轮,小组 6 人中有 5 人或 6 人选择桥路。每组一个二元结局
次要结局 第 31—40 轮的平均通行时间(不含积分);各轮选择桥路的比例;换路次数;收益;任务后对「外部性」的理解(由不知组别的评分者编码)
分析 小组水平的逻辑斯谛回归:陷入 ~ 组别 + 实施地点分层,报告风险差及其 95% 置信区间;双侧 α=0.05\alpha = 0.05

主要结局只设一项。平均通行时间作为最重要的次要结局报告,因为 A 组的积分会改变收益,而收益不能直接比较。

样本量

推导

# Chapter 8 trial. Part 1: the 6-player Braess game of Chapter 3 (links s->v and w->t
# cost 10 per user, v->t and s->w cost 70, bridge v->w 0) played for 40 rounds by simple
# learning BOTS: a MODEL of the hypothesis, not data about people. Control bots see
# every route's time and minimise time; treatment bots also pay congestion points equal
# to the delay they add to others, (n_e - 1) x 10 on each congested link. Primary
# outcome per group: TRAPPED = in at least 8 of rounds 31-40, five or six players took
# the bridge route. Part 2: sample size.
import math, random
from statistics import NormalDist
rng = random.Random(15)
ROUTES = {"U": ("sv", "vt"), "L": ("sw", "wt"), "Z": ("sv", "vw", "wt")}
COST = {"sv": lambda n: 10 * n, "vt": lambda n: 70, "sw": lambda n: 70,
        "wt": lambda n: 10 * n, "vw": lambda n: 0}

def route_cost(others, r, points):
    """Cost of route r for a player when the other five chose 'others'."""
    n = {e: sum(e in ROUTES[o] for o in others) + (e in ROUTES[r]) for e in COST}
    t = sum(COST[e](n[e]) for e in ROUTES[r])
    if points: t += sum((n[e] - 1) * 10 for e in ROUTES[r] if e in ("sv", "wt"))
    return t

def group(points, heed=1.0, rounds=40, beta=0.2, phi=0.7):
    att = [{r: -100.0 + rng.gauss(0, 10) for r in ROUTES} for _ in range(6)]
    uses = [points and rng.random() < heed for _ in range(6)]  # bots counting points
    trap, late = 0, 0
    for t in range(rounds):
        ch = []
        for a in att:
            w = {r: math.exp(beta * (a[r] - max(a.values()))) for r in ROUTES}
            u, acc = rng.random() * sum(w.values()), 0.0
            for r in ROUTES:
                acc += w[r]
                if u <= acc: ch.append(r); break
        for i, a in enumerate(att):             # learn from all routes' costs this round
            others = ch[:i] + ch[i + 1:]
            for r in ROUTES:
                a[r] = phi * a[r] - (1 - phi) * route_cost(others, r, uses[i])
        if t >= rounds - 10:
            trap += ch.count("Z") >= 5; late += ch.count("Z")
    return trap >= 8, late / 60

for pts, heed, lab in ((False, 1, "route times only"),
                       (True, 1, "plus externality points"),
                       (True, 0.5, "points, half the bots ignore them")):
    res = [group(pts, heed) for _ in range(1000)]
    print(f"bots, {lab:33s}: trapped {sum(r[0] for r in res) / 1000:.0%};"
          f" bridge share in rounds 31-40 {sum(r[1] for r in res) / 1000:.0%}")

def n_per_arm(p1, p2, alpha=0.05, power=0.90):
    z = NormalDist().inv_cdf; pb = (p1 + p2) / 2
    num = (z(1 - alpha / 2) * math.sqrt(2 * pb * (1 - pb))
           + z(power) * math.sqrt(p1 * (1 - p1) + p2 * (1 - p2)))
    return math.ceil(num ** 2 / (p1 - p2) ** 2)

print("assumed share trapped: control -> treatment"
      "   analysable GROUPS per arm (90 % power)")
for p1, p2 in ((0.60, 0.30), (0.60, 0.35), (0.70, 0.40), (0.50, 0.25)):
    print(f"                       {p1:.2f} -> {p2:.2f}            {n_per_arm(p1, p2)}")
n = n_per_arm(0.60, 0.30); rec = math.ceil(n / 0.9)
print(f"chosen 0.60 -> 0.30: {n} groups per arm; allowing 10 % incomplete groups,"
      " recruit")
print(f"  {rec} groups per arm = {6 * rec} participants per arm, {12 * rec} in all")
hits = 0
for _ in range(4000):
    a = sum(rng.random() < 0.60 for _ in range(n))
    b = sum(rng.random() < 0.30 for _ in range(n))
    pa, pb = a / n, b / n; pp = (a + b) / (2 * n)
    hits += abs(pa - pb) / math.sqrt(pp * (1 - pp) * 2 / n) > 1.96
print(f"simulation check of power with {n} groups per arm: {hits / 4000:.2f}")
# bots, route times only                 : trapped 77%; bridge share in rounds 31-40 87%
# bots, plus externality points          : trapped 0%; bridge share in rounds 31-40 21%
# bots, points, half the bots ignore them: trapped 4%; bridge share in rounds 31-40 50%
# assumed share trapped: control -> treatment   analysable GROUPS per arm (90 % power)
#                        0.60 -> 0.30            56
#                        0.60 -> 0.35            82
#                        0.70 -> 0.40            56
#                        0.50 -> 0.25            77
# chosen 0.60 -> 0.30: 56 groups per arm; allowing 10 % incomplete groups, recruit
#   63 groups per arm = 378 participants per arm, 756 in all
# simulation check of power with 56 groups per arm: 0.89

程序第一部分不是数据,而是对假说的一个模型:每个「机器人」按以往各路线的成本加权选择(成本低的路线更可能被选,但保留随机性),每轮之后用所有路线的成本更新自己的估计。只看路线时间的机器人中,77% 的小组陷入布雷斯均衡,最后 10 轮有 87% 的选择是桥路,与拉波波特等人发现的多数人最终走桥在方向上一致;扣积分的机器人中没有一个小组陷入,桥路选择降到 21%;即使一半机器人完全忽略积分,陷入的比例也只有 4%。这个差距是一个上限:机器人对积分的反应是设计好的,真实的人可能不理解、不在意或怀疑积分的意义。

所以样本量按保守得多的假设计算:对照组陷入 60%,A 组 30%,双侧 α=0.05\alpha = 0.05,把握度 90%:每组需要 56 个可分析的小组;考虑约 10% 的小组因有人中途退出而不完整,每组招募 63 个小组,即每组 378 人,共 756 人;模拟检验的把握度为 0.89。若真实差距只有 25 个百分点(60% 对 35%),每组需要 82 个小组。这些参数都是假设值,正式实施前应当用预试验的数据核对。

注

为什么以小组为单位。 同一小组内 6 个人的选择互相影响,结局本身就是小组的状态,所以随机化与分析都以小组为单位,样本量按小组数计算;每组只有一个结局值,所以不需要组内相关系数与设计效应。每组只有 6 人,属于第二章所说的有限人数博弈;拉波波特等人 2009 年的实验每组约 18 人 (Rapoport et al. 2009),本方案的规模较小,结论不能直接推广到人数更多的情形。

判定规则

推导

# Decision rule for the Chapter 8 trial, fixed before any data are collected. diff =
# share of TRAPPED groups in the control arm (route times only) minus the treatment arm
# (own extra delay to others + congestion points); POSITIVE diff = fewer trapped groups
# with the treatment. 95 % CI from the pre-specified model (logistic regression on the
# group-level outcome, arm + site stratum, reported as a risk difference). Smallest
# difference of interest: 0.15 (15 percentage points). Fidelity: the treatment screen
# must have been shown as specified in >= 90 % of treatment-arm rounds.
SESOI = 0.15

def verdict(diff, lo, hi, fidelity, time_change):
    if fidelity < 0.90:
        return (f"diff {diff:+.2f}: NOT INTERPRETABLE - screen shown as"
                f" specified"
                f" in only {fidelity:.0%} of rounds")
    if lo >= SESOI:    v = "treatment cuts trapping by >= 15 points"
    elif lo > 0:       v = "treatment cuts trapping, >= 15 points not shown"
    elif hi < 0:       v = "treatment INCREASES trapping"
    elif lo > -SESOI and hi < SESOI: v = "no difference worth acting on"
    else:              v = "unclear: the interval is too wide to decide"
    return f"diff {diff:+.2f} ({lo:+.2f} to {hi:+.2f}): {v}; time {time_change:+.0%}"

for case in ((0.33, 0.17, 0.49, 0.97, -0.09), (0.18, 0.02, 0.34, 0.95, -0.04),
             (0.02, -0.12, 0.14, 0.96, 0.00), (-0.20, -0.36, -0.04, 0.95, 0.05),
             (0.35, 0.19, 0.51, 0.82, -0.10), (0.10, -0.08, 0.28, 0.94, -0.02)):
    print(verdict(*case))
# diff +0.33 (+0.17 to +0.49): treatment cuts trapping by >= 15 points; time -9%
# diff +0.18 (+0.02 to +0.34): treatment cuts trapping, >= 15 points not shown; time -4%
# diff +0.02 (-0.12 to +0.14): no difference worth acting on; time +0%
# diff -0.20 (-0.36 to -0.04): treatment INCREASES trapping; time +5%
# diff +0.35: NOT INTERPRETABLE - screen shown as specified in only 82% of rounds
# diff +0.10 (-0.08 to +0.28): unclear: the interval is too wide to decide; time -2%

示例结果是编造的。A 组界面没有按规定显示的轮次超过 10% 时,无论结果如何,报告都写「不可解释」。只有当陷入比例之差的 95% 置信区间下限不低于 15 个百分点时,报告才写「干预使陷入至少减少了最小关注的幅度」。平均通行时间与收益只报告估计与区间,不用来推翻主要结论。

这项试验不能说明什么

第一,A 组同时得到两样东西:关于外部性的信息与按外部性扣分的价格。阳性结果不能区分是哪一样起了作用;第三章的「外部性可见假说」需要增加一个「只显示、不扣分」的第三组来检验。第二,这是 40 轮、每组 6 人的实验室博弈,网络结构是已知的、刻意构造的布雷斯网络;真实的道路、组织流程与注意力市场的结构未知、人数多得多。结论不能直接外推到任何真实的交通政策、组织管理或平台规则。第三,积分规则使 A 组的收益结构不同,参与者可能为了少扣分而改变行为,这正是干预的意图,但它也意味着 A 组的「好结果」部分来自外在激励,撤去积分后能否维持,本试验不能回答。第四,参与者多为大学生。第五,只用了一组参数。第六,主要结局是小组状态的二元判断,门槛(10 轮中至少 8 轮、6 人中至少 5 人)是事先规定的,换一种门槛结论可能不同,因此次要结局会报告各轮桥路比例的完整曲线。

伦理与利益冲突

这是一项低风险研究:任务是计算机上的路线选择游戏,不涉及任何健康干预,也不收集健康信息,可预见的不适只是收益较少时的挫败感。即便如此,试验仍须经独立伦理委员会审查批准,每位参与者签署知情同意,并在第一个小组入组之前登记方案与分析计划 (Nosek et al. 2018)。任务说明如实告知收益与积分规则;不告知研究假设属于信息保留而非欺骗,任务结束后向所有人说明研究目的。报酬规则事先公布;任何人可随时退出而无需说明理由,并获得已完成部分的报酬,该小组记为不完整。数据匿名保存。无论结果正负,都将发表。

利益披露

作者以 5D Library 的著作(包括本书)作为其社群课程、升维牌卡(5D Ascension Cards,Dimensional Shift Card Decks)与 Yafu AI 知识库的方法论基础,因此与本书主题有商业利益关系。作者还创办并经营 Asia Influencer X 与 Yafu Global,业务涉及多平台流量分发与数字营销矩阵,与本书主题有直接的商业利益。本书不推销任何课程、牌卡或服务,不是投资或经营建议,书中不使用上述企业的客户数据;第八章的试验设计为由独立于作者企业的研究人员实施和分析,不从作者的社群、学员或客户(包括上述企业的客户)中招募;书中所有代码与研究方案均公开,供任何人检验。

参考文献

Gisches, Eyran J., and Amnon Rapoport. 2012. “Degrading Network Capacity May Improve Performance: Private Versus Public Monitoring in the Braess Paradox.” Theory and Decision 73 (2): 267–93. https://doi.org/10.1007/s11238-010-9237-0.
Little, John D. C. 1961. “A Proof for the Queuing Formula: L=λWL = \lambda W.” Operations Research 9 (3): 383–87. https://doi.org/10.1287/opre.9.3.383.
Morgan, John, Henrik Orzen, and Martin Sefton. 2009. “Network Architecture and Traffic Flows: Experiments on the Pigou–Knight–Downs and Braess Paradoxes.” Games and Economic Behavior 66 (1): 348–72.
Nosek, Brian A., Charles R. Ebersole, Alexander C. DeHaven, and David T. Mellor. 2018. “The Preregistration Revolution.” Proceedings of the National Academy of Sciences 115 (11): 2600–2606. https://doi.org/10.1073/pnas.1708274114.
Rapoport, Amnon, Tamar Kugler, Subhasish Dugar, and Eyran J. Gisches. 2009. “Choice of Routes in Congested Traffic Networks: Experimental Tests of the Braess Paradox.” Games and Economic Behavior 65 (2): 538–71. https://doi.org/10.1016/j.geb.2008.02.007.
Selten, Reinhard, Thorsten Chmura, Thomas Pitz, Sebastian Kube, and Michael Schreckenberg. 2007. “Commuters Route Choice Behaviour.” Games and Economic Behavior 58 (2): 394–406. https://doi.org/10.1016/j.geb.2006.03.012.

Fan, Y. (2026). Algorithmic Game Theory, Braess's Paradox and the Price of Anarchy in Traffic and Attention Networks: From Wardrop Equilibria and Selfish Routing to Platform Congestion, Read Beside Classical Chinese Hydraulic Engineering [算法博弈论、布雷斯悖论与流量分配的无秩序代价] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23782094 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).