关于「预测误差:心智模拟为何得不到下修」,《心智模拟与行为兑现》给出的定量结论是什么?
结论:第一章用学习规则说明,完全可预期的想象学会后不再产生正向误差,真正的问题是缺少负向误差:示意轨迹被模拟抬高到比真实值高出 0.600,且永不下修。
数据口径:书中参数为示意;全部计算由书中程序生成;本书不是心理、临床或婚恋咨询,也不是个人化经营建议。 第一章用学习规则说明,完全可预期的想象学会后不再产生正向误差,真正的问题是缺少负向误差:示意轨迹被模拟抬高到比真实值高出 0.600,且永不下修。
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: 0.600
| A. value estimate V after t trials (alpha 0.15, V0 0.70, p_true 0.40) | |
| trial real R sim S echo B | |
| 0 | 0.700 0.700 0.700 |
| 5 | 0.666 0.867 0.722 |
| 10 | 0.538 0.941 0.745 |
| 20 | 0.466 0.988 0.790 |
| 40 | 0.372 1.000 0.880 |
| 80 | 0.507 1.000 0.999 |
| B. negative prediction errors in 80 trials | |
| real R negative 48 mean delta 1-10 -0.1081 | 71-80 +0.0906 |
| sim S negative | 0 mean delta 1-10 +0.1606 71-80 +0.0000 |
| echo B negative | 0 mean delta 1-10 +0.0300 71-80 +0.0032 |
| C. final gap sim S minus truth +0.600; real R error +0.107 |
查看完整验算程序
# Chapter 1. p11: Rescorla-Wagner value tracks, real feedback vs simulation.
# Model: V <- V + alpha * (r - V), delta = r - V is the prediction error.
# Track R (real): outcomes r are drawn from Bernoulli(p_true) with a fixed
# seed, so the estimate meets negative errors whenever r = 0.
# Track S (simulated): the imagined outcome is the wished outcome r = 1
# on every trial; nothing in the stream can ever fall short of it.
# Track B (biased echo): the imagined outcome is the current estimate
# plus a small optimistic bias b, capped at 1.
# All parameters are illustrative (alpha 0.15, p_true 0.40, V0 0.70).
import math
import random
ALPHA = 0.15
P_TRUE = 0.40
V0 = 0.70
BIAS = 0.03
TRIALS = 80
def run_real(rng):
v, vs, ds = V0, [V0], []
for _ in range(TRIALS):
r = 1.0 if rng.random() < P_TRUE else 0.0
d = r - v
v += ALPHA * d
vs.append(v)
ds.append(d)
return vs, ds
def run_sim(echo):
v, vs, ds = V0, [V0], []
for _ in range(TRIALS):
r = min(1.0, v + BIAS) if echo else 1.0
d = r - v
v += ALPHA * d
vs.append(v)
ds.append(d)
return vs, ds
def main():
rng = random.Random(43011)
vr, dr = run_real(rng)
vs, ds = run_sim(False)
vb, db = run_sim(True)
print("A. value estimate V after t trials (alpha {:.2f}, V0 {:.2f},"
" p_true {:.2f})".format(ALPHA, V0, P_TRUE))
print(" trial real R sim S echo B")
for t in (0, 5, 10, 20, 40, 80):
print(" {:>5} {:.3f} {:.3f} {:.3f}".format(
t, vr[t], vs[t], vb[t]))
print("B. negative prediction errors in {} trials".format(TRIALS))
for name, ds_ in (("real R", dr), ("sim S", ds), ("echo B", db)):
neg = 0
for d in ds_:
if d < 0:
neg += 1
early = math.fsum(ds_[:10]) / 10
late = math.fsum(ds_[70:]) / 10
print(" {:<7} negative {:>2} mean delta 1-10 {:+.4f}"
" 71-80 {:+.4f}".format(name, neg, early, late))
gap = vs[TRIALS] - P_TRUE
print("C. final gap sim S minus truth {:+.3f}; real R error {:+.3f}"
.format(gap, vr[TRIALS] - P_TRUE))
if __name__ == "__main__":
main()唯一数据源与引用
范雅福(Fan, Y.). (2026). 《心智模拟亲密与具身行为兑现的神经分野:多巴胺投射幻象与真实契约代价母典》第 1 章「预测误差:心智模拟为何得不到下修」(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-430 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026430,
author = {Fan, Yafu},
title = {Mental Simulation versus Behavioural Enactment in Intimacy: Dopaminergic Fantasy Loops, Oxytocin Bonding Costs and Commitment Asymmetries},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}