关于「适存胜过真实:一比特知觉的复制子竞赛」,《效用、真实与双账本》给出的定量结论是什么?
结论:第一、二章用复制子动力学与 Moran 过程复现「适存胜过真实」:一比特知觉、单峰收益下,适存编码策略的份额从起点到第 60 个时间单位升到 0.7712;但目标一旦从一个增加到两个,真实编码就在 0.703 的随机目标组上不输,收益模型设错或存在悬崖时,能如实刻画边界的编码胜出。这是本书的蓝图修正:效用胜过真实只在写明的有界模型内成立。
数据口径:书中参数与数据均为示意或合成,全部计算由书中程序生成;本书不是医疗建议,不提供任何临床心理服务,不是个人投资建议,不承诺任何收益;功能性信念只用于自己的行动框架,不用于欺骗他人。 第六章建立双账本:行动框架可按效用编码,概率账以 Brier 分数校准,硬边界账不让步;框架改写概率账时,示意破产比例从 0.0925 升到 0.7055。
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: 0.7712
thresholds: truth x > 0.27, fitness g(x) > 0.50 solo payoff: truth 0.5480, fitness 0.6895 payoff matrix (row meets column): T vs T 0.5149 T vs F 0.4645 F vs T 0.6137 F vs F 0.5567 replicator, fitness share from 0.01 (Euler step 0.1): time 10.0 fitness share 0.0263 time 20.0 fitness share 0.0670 time 30.0 fitness share 0.1602 time 40.0 fitness share 0.3347 time 60.0 fitness share 0.7712 time 400.0 fitness share 1.0000
查看完整验算程序
# Chapter 1. p11: fitness-coded vs truth-coded perception, replicator run
# Model after Mark, Marion and Hoffman (2010): a resource quantity x is
# uniform on [0, 1]; its fitness payoff g(x) is single-peaked (too little
# or too much of the resource is worth less than the middle). Each
# strategy sees one bit per territory. TRUTH colours a territory "high"
# when x > t (an order-preserving split of x); FITNESS colours it "high"
# when g(x) > s. Both thresholds are tuned for a solo choice among three
# territories. In a contest, three territories are drawn, a random first
# mover takes the one it sees as "high", and the second mover takes its
# preferred one of the two left. All parameters are illustrative.
import math
import random
PEAK, WIDTH = 0.5, 0.18
DRAWS = 60000
def g(x):
return math.exp(-0.5 * ((x - PEAK) / WIDTH) ** 2)
def pick(colours, rng):
best = [i for i, c in enumerate(colours) if c]
pool = best if best else list(range(len(colours)))
return pool[rng.randrange(len(pool))]
def solo_payoff(rule, cut, terr):
rng = random.Random(42711)
vals = []
for xs in terr:
cols = [rule(x) > cut for x in xs]
vals.append(g(xs[pick(cols, rng)]))
return math.fsum(vals) / len(vals)
def tune(rule, terr):
grid = [i / 100 for i in range(5, 96)]
return max(grid, key=lambda c: (solo_payoff(rule, c, terr), -c))
rng0 = random.Random(42701)
TERR = [[rng0.random() for _ in range(3)] for _ in range(DRAWS)]
tune_set = TERR[:8000]
t_cut = tune(lambda x: x, tune_set)
s_cut = tune(g, tune_set)
RULES = {"T": (lambda x: x, t_cut), "F": (g, s_cut)}
def contest(a, b):
rng = random.Random(42702)
pay = []
for n, xs in enumerate(TERR):
a_first = n % 2 == 0
first, second = (a, b) if a_first else (b, a)
r1, c1 = RULES[first]
i = pick([r1(x) > c1 for x in xs], rng)
rest = [x for k, x in enumerate(xs) if k != i]
r2, c2 = RULES[second]
j = pick([r2(x) > c2 for x in rest], rng)
pay.append(g(xs[i]) if a_first else g(rest[j]))
return math.fsum(pay) / len(pay)
M = {(a, b): contest(a, b) for a in "TF" for b in "TF"}
print("thresholds: truth x > %.2f, fitness g(x) > %.2f" % (t_cut, s_cut))
print("solo payoff: truth %.4f, fitness %.4f" % (
solo_payoff(lambda x: x, t_cut, TERR), solo_payoff(g, s_cut, TERR)))
print("payoff matrix (row meets column):")
for a in "TF":
print(" %s vs T %.4f %s vs F %.4f" % (a, M[(a, "T")], a, M[(a, "F")]))
xf, dt = 0.01, 0.1
print("replicator, fitness share from 0.01 (Euler step 0.1):")
for step in range(1, 4001):
pf = xf * M[("F", "F")] + (1 - xf) * M[("F", "T")]
pt = xf * M[("T", "F")] + (1 - xf) * M[("T", "T")]
mean = xf * pf + (1 - xf) * pt
xf += dt * xf * (pf - mean)
if step in (100, 200, 300, 400, 600, 4000):
print(" time %5.1f fitness share %.4f" % (step * dt, xf))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《效用优先于字面真实:演化适存度博弈、功能性信念重塑与物理硬边界母典》第 1 章「适存胜过真实:一比特知觉的复制子竞赛」(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-427 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026427,
author = {Fan, Yafu},
title = {When Utility Outperforms Literal Truth: Evolutionary Fitness-Payoff Games, Functional Belief Engineering and Physical Guardrails},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}