因果发现与归纳问题

Causal Discovery and The Problem of Induction

归纳问题是科学哲学与认识论中的核心问题之一。它的目的是寻找从已知预测未知、从有限经验推广到一般规律的归纳推断的合理性的基础。自从休谟提出归纳怀疑论的现代表述以来,尽管人们尝试了许多方式试图对其进行回应,但它们都没有得到普遍接受。本文论证的观点是,在机器学习领域所通行的验证因果发现算法的推理模式提供了一种回应归纳怀疑论的特殊方法。这种方法不但具有与传统的科学哲学确证理论不同的特征,并且还能够揭示出归纳推断问题的更为精细的结构。

The problem of induction is a central issue in both philosophy of science and epistemology.It is about the validation of inductive inference from observed data to unobserved conclusion.Since Hume raised the problem in its modern form,there were many attempts to respond but none is widely accepted.This paper focuses on the problem of the machine learning discovery of causal relations among a finite set of variables from purely observed data.It is argued that the way machine learning researchers validate automatic causal discovery algorithms provides a novel approach toward the philosophical problem of inductive skepticism.This new approach shows that there are additional fine-grained structures of inductive inferences that is more appropriate to characterize scientific inferences than the traditional understanding of induction.