Observational Causal Induction with Small Samples

Abstract

Causal induction forms the basis for adaptation of animals. It enables diagnosis of the past, judgment in the present, and prediction for the future in uncertain environments. In AI, it is argued that efficient causal inference supports human-like quick learning and decision-making with small amount of data. The purpose of this study is to elucidate how humans judge the causal relationship when given small data. We assume that causal induction has two stages, observational and interventional, according to Hattori and Oaksford. In our experiment, we controlled a factor (symmetry between occurrence and non-occurrence of the causal events) that has been shown to manipulate the cognitive distinction between the two stages. We compared the models that correspond to the two stages, pARIs/DFH and delta-P. It was shown that pARIs, the probability of biconditionals and equivalent to Jaccard index, describes the result best in both conditions.