关于「奖赏预测误差:时冷时热关系的时序差分学习小模型」,《真心的信号与时冷时热的陷阱》给出的定量结论是什么?

结论:平均温暖的排序是 S1、S2、S3、S4:0.800、0.400、0.372、0.244。冲劲的排序几乎完全相反:S1 与 S2 都是 0.000,S3 是 0.245,S4 是 0.193;最大的单次冲劲出现在 S3(0.917)与 S4(0.887)。如果学习者用「感觉有多强烈」来判断谁值得留下,他会选中给得最少的两位;如果用平均温暖来判断,他会选中 S1。

数据口径:书中参数均为示意值;全部计算由书中程序生成;本书不是心理咨询,不宣称任何效果,也不教授任何操控手段。 第二章把每次互动记入似然比账本,示意参数下一次失约的分量相当于 2.52 次兑现,沃尔德序贯检验平均比固定观察期节省 0.459 的互动。

English version

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: 0.000, 0.193, 0.244, 0.245, 0.372, 0.400

partner mean reward rush/round biggest rush leaves at
S1 steady 0.80.800 0.000 0.009 stays
S2 steady 0.40.400 0.000 0.005 round 21
S3 random 0.40.372 0.245 0.917 round 25
S4 flood+cycle0.244 0.193 0.887 round 29
rush ranks S3 and S4 on top; mean reward ranks them last
查看完整验算程序
# Chapter 4. p42: felt intensity is not value. A delta-rule learner
# (learning rate 0.2, start 0) meets four partners for 200 rounds:
# S1 steady warm 0.8 each round; S2 steady lukewarm 0.4; S3 random,
# warm (1) with probability 0.4; S4 a 20-round flood of warmth, then a
# cycle of 6 cold rounds and 2 warm ones. "Rush" = mean positive error
# per round, rounds 21 to 200. The average rule leaves once the mean
# reward of the last 12 rounds falls below 0.5 (checked from round 21).
# Illustrative values; S3 uses a fixed seed.
import math
import random

ALPHA, N, WIN, FLOOR = 0.2, 200, 12, 0.5
rng = random.Random(20261112)


def s1(t):
    return 0.8


def s2(t):
    return 0.4


def s3(t):
    return 1.0 if rng.random() < 0.4 else 0.0


def s4(t):
    if t < 20:
        return 1.0
    return 1.0 if (t - 20) % 8 >= 6 else 0.0


print("partner          mean reward  rush/round  biggest rush  leaves at")
for lab, f in (("S1 steady 0.8", s1), ("S2 steady 0.4", s2),
               ("S3 random 0.4", s3), ("S4 flood+cycle", s4)):
    v, rs, pos, top, leave = 0.0, [], [], 0.0, None
    for t in range(N):
        r = f(t)
        d = r - v
        v += ALPHA * d
        rs.append(r)
        if t >= 20:
            pos.append(max(d, 0.0))
            top = max(top, d)
            recent = math.fsum(rs[-WIN:]) / WIN
            if leave is None and recent < FLOOR:
                leave = t + 1
    mr = math.fsum(rs[20:]) / (N - 20)
    lv = "round {:3d}".format(leave) if leave else "  stays"
    print("{:15s}  {:11.3f}  {:10.3f}  {:12.3f}  {}".format(
        lab, mr, math.fsum(pos) / len(pos), top, lv))
print("rush ranks S3 and S4 on top; mean reward ranks them last")

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《亲密关系早期高代价信号甄别、间歇性强化多巴胺成瘾陷阱与对等底线停时博弈》第 4 章「奖赏预测误差:时冷时热关系的时序差分学习小模型」(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-181 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026181,
  author    = {Fan, Yafu},
  title     = {Costly-Signal Screening in Early Intimacy, Intermittent-Reinforcement Dopamine Traps and Reciprocal-Floor Stopping Games},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论