Skip to main content
QUICK REVIEW

[论文解读] A defense of Columbo (and of the use of Bayesian inference in forensics): A multilevel introduction to probabilistic reasoning

G. D’Agostini|arXiv (Cornell University)|Mar 10, 2010
Bayesian Modeling and Causal Inference参考文献 5被引用 5
一句话总结

本文通過分析一集著名的《科隆博》劇集,為法醫推理中使用貝葉斯推斷進行辯護,證明這種概率推理方法是有效的。利用包含不確定證詞的多層次貝葉斯網絡模型,本文表明偵探透過誘使兇手指出正確攝影機的作法,結合先驗信念與目擊者可信度更新後,可形成強而一致的證據,從而反駁了法庭上存在「概率陷阱」的說法。

ABSTRACT

Triggered by a recent interesting New Scientist article on the too frequent incorrect use of probabilistic evidence in courts, I introduce the basic concepts of probabilistic inference with a toy model, and discuss several important issues that need to be understood in order to extend the basic reasoning to real life cases. In particular, I emphasize the often neglected point that degrees of beliefs are updated not by `bare facts' alone, but by all available information pertaining to them, including how they have been acquired. In this light I show that, contrary to what claimed in that article, there was no "probabilistic pitfall" in the Columbo's episode pointed as example of "bad mathematics" yielding "rough justice". Instead, such a criticism could have a `negative reaction' to the article itself and to the use of Bayesian reasoning in courts, as well as in all other places in which probabilities need to be assessed and decisions need to be made. Anyway, besides introductory/recreational aspects, the paper touches important questions, like: role and evaluation of priors; subjective evaluation of Bayes factors; role and limits of intuition; `weights of evidence' and `intensities of beliefs' (following Peirce) and `judgments leaning' (here introduced), including their uncertainties and combinations; role of relative frequencies to assess and express beliefs; pitfalls due to `standard' statistical education; weight of evidences mediated by testimonies. A small introduction to Bayesian networks, based on the same toy model (complicated by the possibility of incorrect testimonies) and implemented using Hugin software, is also provided, to stress the importance of formal, computer aided probabilistic reasoning.

研究动机与目标

  • 反駁『科隆博』劇集《負面反應》一集中的場景在法庭證據評估中存在『概率陷阱』的說法。
  • 證明當完整資訊(包括證據收集背景)被納入考量時,正確應用貝葉斯推斷可避免常見的概率推理錯誤。
  • 強調在法醫決策中整合先驗信念、目擊者可信度與證據結構的重要性。
  • 提倡使用貝葉斯網絡作為正式工具,以管理不確定性,避免過度依賴直覺或頻率學派的誤解。
  • 顯示即使個別目擊者可信度不確定,多組一致的證詞仍可增強對其真實性的信念。

提出的方法

  • 使用一個包含12顆黑球與1顆白球的盒子模型,模擬法醫證據收集與目擊者證詞。
  • 應用貝葉斯定理,根據證據與證詞更新對真實狀態(例如:哪台攝影機被使用)的信念程度。
  • 引入在Hugin軟體中實現的貝葉斯網絡,以建模盒子組成、目擊者報告與說謊機率之間的依賴關係。
  • 計算有效貝葉斯因子,以量化證據支持或反對假說的強度,使用公式 $\tilde{O}_{B_1}(B_T,I) = \frac{1}{P(W|B_2) + P(B|B_2)/\lambda}$。
  • 將目擊者可信度建模為潛在變數,其說謊的先驗機率為(例如)$\lambda = 1/5$,並根據報告的一致性更新信念。
  • 透過迭代網絡更新模擬順序性證據:首先兩名目擊者稱為『白球』,第三名稱為『黑球』,隨後直接觀察到一顆黑球。

实验结果

研究问题

  • RQ1科隆博偵探誘使兇手指出正確攝影機的作法,是否為合適的概率推理應用,還是構成了『概率陷阱』?
  • RQ2多組一致或衝突的證詞,如何影響對自然真實狀態的後驗信念?
  • RQ3先驗信念與證據收集方法在法醫案件中對最終機率評估的影響程度為何?
  • RQ4貝葉斯網絡能否形式化並改善法律與科學情境中對不確定證據的直覺推理?
  • RQ5當證詞不確定或衝突時,貝葉斯因子在量化證據強度方面扮演何種角色?

主要发现

  • 在考慮情境與兇手的行為並非隨機而是對科隆博提問的回應後,正確攝影機被識別出的機率從1/12上升至75.7%。
  • 在兩名目擊者報告『白球』、一名報告『黑球』後,盒子為$B_1$(白球)的機率降至75.7%,貝葉斯因子為3.1,顯示對$B_1$有中等程度的反證。
  • 在兩名目擊者報告一致後,相信他們說真話的機率從85%上升至95%,顯示一致性可提升可信度。
  • 在第三名目擊者與前兩名衝突後,對前兩名目擊者說真話的信念下降至83%,而對第三名目擊者的信任度也降低至76%,顯示多數報告會獲得更高權重。
  • 在直接觀察到一顆黑球後,未來抽到白球的機率降至7.7%,且相信前兩名目擊者說謊的機率上升至70.6%,而第三名目擊者說謊的機率則降至1.6%。
  • 網絡模型確認,定性直覺與定量貝葉斯結果一致,但唯有計算才能精確得出不確定性下信念更新的估計值。

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。