因果强度推理的贝叶斯模型综述

A Review of the Bayesian Models of Causal Strength Inference

因果强度推理的贝叶斯模型以因果协变关系、贝叶斯定理和因果图形模型为基础,借助蒙特卡洛算法实现模型对被试因果强度估计的预测。通过选用不同的先验分布、似然函数、蒙特卡洛取样方法和基于后验分布的预测方法,因果强度推理的贝叶斯模型可以精确预测很多以往模型不能解释的实验结果,并对被试的推理过程提出有价值的见解,还可推广到对因果结构判断的预测和对多原因(结果)交互作用的解释;但需要在如何选择先验分布和似然函数、解释被试如何表征与作答、扩大单个模型解释范围和简化计算等方面继续完善。

On the basis of the essence of Bayesian statistics and the progress of computational technology, Bayesian models of causal strength inference have got rapid development in the last two decades. These models draw post distribute out of the combination of prior distribute and observationdata, and then are used to make prediction based on the post distribute. Different models comprise various pri- or distribute （ assign, uniform, sparse and strong, experiment, et al）, likelihood function （ Noisy - Or, Noisy - AND - NOT, Noisy - Logi- cal et al）, and methods that are used to make prediction （compare different post distribute, compute mean value of post distribute et al）. The advantages of Bayesian models include representing the impact of uncertainty by integrating the bottom - up and top - down ap- proach, having a great insight into human participants reasoning process, and playing better to predict participants＇ performance than other models. These models need to improve on how to choose appropriate prior distribute, how to explain participants＇ various operation on different conditions, and how to decrease computational intractability. The present paper is a brief introduction of Bayesian model＇ s theoretical basis, mathematical compositions and practical application in causal strength inference.