关于「新颖着法率:首个历史新着、数据库膨胀与背谱检验」,《围棋 AI 后的人类决策质量》给出的定量结论是什么?
结论:第三章说明新颖指数随历史库膨胀而机械漂移,并检验插入 AI 对局的反背谱检验:照抄棋谱会被完全识破,学会 AI 的开局分布却仍保留 0.860 的增益。
数据口径:书中除论文数字外的参数均为示意,全部计算由书中程序生成;本书是非个人化经营建议,不针对任何个人决定。 第四章用香农熵度量定式集中度,示意角部局面的策略熵从 1.608 升到 2.360 比特,从众学习模型显示外部评估者让群体提早约 12 年离开局部最优。
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: 0.860
| mean NI of human games 2017 to 2021 | ||||
| no AI influence at all: 592.16 | ||||
| world | without insert | with insert | change | gain kept |
| C | 592.59 | 590.13 | -2.46 | 0.000 |
| M | 594.59 | 594.25 | -0.34 | 0.860 |
| I | 594.60 | 594.58 | -0.02 | 0.991 |
查看完整验算程序
# Chapter 3. p33: the inserted-AI-games test for memorised novelty
# Model (illustrative, built on p32): from 2017 the AI-influenced human
# games either replay the first 12 moves of one of the AI's own games
# (world C, copying lines), or draw from the AI's opening distribution
# (world M, rotation 3, s = 1.6, the law AI-vs-AI games follow), or
# explore a region of their own (world I: rotation 6, s = 1.6).
# The test of Shin et al. (2023) is mimicked: 600 AI-vs-AI games are
# inserted into the history at the end of 2016, then the Novelty Index
# of human games in 2017 to 2021 is recomputed. Memorised novelty
# should mostly vanish; self-made novelty should survive. Seeded.
import bisect
import math
import random
import statistics
SEED = 20261112
DEPTH = 30
CANDS = 10
S_HUMAN = 2.2
S_AI = 1.6
AI_ROT = 3
INSERTED = 600
MAX_MOVE = 602
def cumulative(s):
w = [1.0 / (i + 1) ** s for i in range(CANDS)]
total = math.fsum(w)
return [math.fsum(w[:k + 1]) / total for k in range(CANDS)]
def pick(rng, cum):
return min(bisect.bisect_left(cum, rng.random()), CANDS - 1)
def play(rng, cum_h, cum_ai, rot):
seq = []
for k in range(DEPTH):
if rot and k >= 2:
seq.append((pick(rng, cum_ai) + rot) % CANDS)
else:
seq.append(pick(rng, cum_h))
return tuple(seq)
def novel_move(seq, seen):
first = None
for k in range(1, DEPTH + 1):
p = seq[:k]
if first is None and p not in seen:
first = k
seen.add(p)
return first if first is not None else DEPTH + 1
def share(year):
return {2017: 0.3, 2018: 0.5}.get(year, 0.7 if year >= 2019 else 0.0)
def post_ai_ni(human_rot, insert):
rng = random.Random(20261112)
ins_rng = random.Random(20261113)
cum_h, cum_ai = cumulative(S_HUMAN), cumulative(S_AI)
ai_games = [play(ins_rng, cum_h, cum_ai, AI_ROT)
for _ in range(INSERTED)]
seen, post = set(), []
for year in range(1980, 2022):
if year == 2017 and insert:
for g in ai_games:
novel_move(g, seen)
for _ in range(60 + 6 * (year - 1980)):
rot = human_rot if rng.random() < share(year) else 0
seq = play(rng, cum_h, cum_ai, rot % CANDS)
if rot == CANDS:
seq = ai_games[rng.randrange(INSERTED)][:12] + seq[12:]
m = novel_move(seq, seen)
if year >= 2017:
post.append(MAX_MOVE - m)
return statistics.fmean(post)
if __name__ == "__main__":
base = post_ai_ni(0, False)
print("mean NI of human games 2017 to 2021")
print("no AI influence at all: {:.2f}".format(base))
print("world without insert with insert change gain kept")
for name, rot in (("C", CANDS), ("M", AI_ROT), ("I", 6)):
a = post_ai_ni(rot, False)
b = post_ai_ni(rot, True)
kept = max(0.0, b - base) / (a - base)
print(" {} {:.2f} {:.2f} {:+.2f} {:.3f}"
.format(name, a, b, b - a, kept))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《围棋人工智能后人类决策质量跃迁、新颖着法率与棋经十三篇策略熵母典》第 3 章「新颖着法率:首个历史新着、数据库膨胀与背谱检验」(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-230 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026230,
author = {Fan, Yafu},
title = {Post-AI Leaps in Human Go Decision Quality, Novel-Move Rates and Strategic Entropy in the Thirteen-Chapter Classic of Go},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}