人工智能中因果定义的探索

The analysis of causality in artificial intelligence

基于统计学的视角,珀尔等人构建了一套能够将预测、干预和反事实算法化的因果语言,通过因果模型表征因果知识,进而给出因果关系的形式化、数学式的定义。HP定义是一种应用性定义,在一定程度上符合因果定义的基本要求:时序性、相关性和不间断性。HP定义可为人工智能领域表征因果知识提供方法,为确立实际原因提供一般性原则。但是,在一定模型中成立的实际因果关系须依赖于语境,模型的选择具有主观性,而并不是在普遍意义上成立的因果关系。在构建模型的过程中,面临着变量是否合理的问题和非递归性因果关系表征等困境。

Based on statistics,Pearl constructed a causal language that could algorithmize prediction,intervention and counterfactual,and represented causal knowledge by causal models,then gave the formal,mathematical definition of causality.The HP definition is an applied definition,which can somehow satisfy the basic demands of causal definition:time-sequence,relevance and uninterrupted.It can present a formal representation of causal knowledge and a principled way of determining actual causes in economics,social science and artificial intelligence.However,the actual causal relationship that holds in a certain model depends on the context,rather than in a general sense.The choice of the model is subjective,so for constructing appropriate models,we are faced with questions about whether variables are reasonable and how non-recursive causation is represented.