关于「拆墙提问序列:示意参数下的模拟最优与敏感度」,《先验消解与反逆火沟通》给出的定量结论是什么?
结论:第六章把追问、肯定、证据与思想实验写成马尔可夫决策过程,示意参数下最难起点的期望收益 10.555,参数扰动两千次,完整路线不变的比例为 0.8270,沿用基准策略的平均相对遗憾只有 0.0003,所以只称模拟最优。
数据口径:书中参数为示意,全部计算由书中程序生成;本书非心理咨询、不提供任何临床服务、非个人化经营建议,例子只用商品与消费认知,不涉及政治、宗教与族群议题。 第六章把追问、肯定、证据与思想实验写成马尔可夫决策过程,示意参数下最难起点的期望收益 10.555,参数扰动两千次,完整路线不变的比例为 0.8270,沿用基准策略的平均相对遗憾只有 0.0003,所以只称模拟最优。
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: 0.0003, 0.8270
| baseline path from D | 2, O=0: AAEEE |
| first action over 2,000 draws: | |
| A | 1.0000 |
| most frequent likely paths: | |
| AAEEE | 0.8270 |
| AAAAQ | 0.1365 |
| AAAAE | 0.0360 |
| AAQEE | 0.0005 |
| relative regret of the baseline policy: | |
| mean 0.0003 median 0.0000 | 95th pct 0.0019 max 0.0081 |
| R first enters the policy at r_gain | 5.5 |
查看完整验算程序
# Chapter 6. p62: how fragile is the "optimal" sequence?
# Same decision process as p61. Every probability and gain parameter is
# multiplied by an independent factor drawn from U(0.75, 1.25) (seeded,
# 2,000 draws; probabilities clipped to [0, 1]). For each draw: the
# optimal first action and most likely path from D=2, O=0, and the regret
# of keeping the baseline policy instead of re-optimizing. Last, the gain
# a thought experiment needs before it ever enters the policy.
import math
import random
H, PEN = 5, 5.0
W = (1.0, 0.5, 0.15) # discount weight by defence level
PAR = {"q_open": 0.5, "q_open_hi": 0.25, "q_def": 0.10,
"a_down": 0.6, "e_def": 0.25, "e_leave": 0.15,
"r_def": 0.15, "r_open": 0.3, "r_leave": 0.20,
"e_gain": 6.0, "r_gain": 3.0, "q_gain": 1.0}
ACTS = "QAER"
def model(d, o, a, p):
"""Return (reward, [(prob, next_state)]), next_state None = left."""
up = lambda x, k: min(2, x + k)
if a == "Q":
po = p["q_open"] if d <= 1 else p["q_open_hi"]
pd = p["q_def"]
out = [(po * pd, (up(d, 1), up(o, 1))),
(po * (1 - pd), (d, up(o, 1))),
((1 - po) * pd, (up(d, 1), o)),
((1 - po) * (1 - pd), (d, o))]
return p["q_gain"] * W[d], out
if a == "A":
pa = p["a_down"] if d > 0 else 0.0
return 0.0, [(pa, (max(0, d - 1), o)), (1 - pa, (d, o))]
if a == "E":
pl = p["e_leave"] if d == 2 else 0.0
pd = p["e_def"] * (1 - 0.3 * o)
rew = (1 - pl) * p["e_gain"] * W[d] * (1 + 0.5 * o)
return rew, [(pl, None), ((1 - pl) * pd, (up(d, 1), o)),
((1 - pl) * (1 - pd), (d, o))]
pl = p["r_leave"] if d == 2 else 0.0
pd, po = p["r_def"], p["r_open"]
rew = (1 - pl) * p["r_gain"] * o * W[d]
out = [(pl, None)]
for dd, qd in ((1, pd), (0, 1 - pd)):
for oo, qo in ((1, po), (0, 1 - po)):
out.append(((1 - pl) * qd * qo, (up(d, dd), up(o, oo))))
return rew, out
def solve(p, horizon=H):
states = [(d, o) for d in range(3) for o in range(3)]
v = {s: 0.0 for s in states}
pol = []
for _ in range(horizon):
nv, step = {}, {}
for s in states:
best = None
for a in ACTS:
r, out = model(s[0], s[1], a, p)
terms = [r] + [q * (-PEN if ns is None else v[ns])
for q, ns in out]
val = math.fsum(terms)
if best is None or val > best[0] + 1e-12:
best = (val, a)
nv[s], step[s] = best
v = nv
pol.insert(0, step)
return v, pol
def likely_path(pol, p, s=(2, 0)):
seq = []
for t in range(len(pol)):
a = pol[t][s]
seq.append(a)
out = model(s[0], s[1], a, p)[1]
q, ns = max(out, key=lambda x: (x[0], x[1] is not None))
if ns is None:
break
s = ns
return "".join(seq)
def evaluate(pol, p, horizon=H):
"""Value of a fixed (time-indexed) policy under parameters p."""
states = [(d, o) for d in range(3) for o in range(3)]
v = {s: 0.0 for s in states}
for t in range(horizon - 1, -1, -1):
nv = {}
for s in states:
r, out = model(s[0], s[1], pol[t][s], p)
nv[s] = math.fsum([r] + [q * (-PEN if ns is None else v[ns])
for q, ns in out])
v = nv
return v
if __name__ == "__main__":
rng = random.Random(38762)
v0, pol0 = solve(PAR)
base_path = likely_path(pol0, PAR)
first, paths, regrets = {}, {}, []
draws = 2000
for _ in range(draws):
p = {}
for k, x in sorted(PAR.items()):
y = x * rng.uniform(0.75, 1.25)
p[k] = min(1.0, y) if not k.endswith("gain") else y
v, pol = solve(p)
a = pol[0][(2, 0)]
first[a] = first.get(a, 0) + 1
lp = likely_path(pol, p)
paths[lp] = paths.get(lp, 0) + 1
vb = evaluate(pol0, p)[(2, 0)]
regrets.append((v[(2, 0)] - vb) / v[(2, 0)])
print("baseline path from D=2, O=0: {}".format(base_path))
print("first action over {:,} draws:".format(draws))
for a in sorted(first):
print(" {} {:.4f}".format(a, first[a] / draws))
top = sorted(paths.items(), key=lambda kv: (-kv[1], kv[0]))[:4]
print("most frequent likely paths:")
for pth, n in top:
print(" {:<6s} {:.4f}".format(pth, n / draws))
regrets.sort()
print("relative regret of the baseline policy:")
print(" mean {:.4f} median {:.4f} 95th pct {:.4f} max {:.4f}"
.format(math.fsum(regrets) / draws, regrets[draws // 2],
regrets[int(0.95 * draws)], regrets[-1]))
for g10 in range(30, 130, 5):
p = dict(PAR)
p["r_gain"] = g10 / 10.0
pol = solve(p)[1]
used = [st[s] for st in pol for s in st if st[s] == "R"]
if used:
print("R first enters the policy at r_gain = {:.1f}".format(
g10 / 10.0))
break
else:
print("R never enters the policy for r_gain up to 12.5")唯一数据源与引用
范雅福(Fan, Y.). (2026). 《顽固先验偏见消解、解释深度错觉破除与反逆火效应沟通母典》第 6 章「拆墙提问序列:示意参数下的模拟最优与敏感度」(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-387 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026387,
author = {Fan, Yafu},
title = {Dissolving Entrenched Priors, Shattering the Illusion of Explanatory Depth and Anti-Backfire Persuasion Mechanics},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}