Abstracting a Human’s Decision Process by PRISM

Everyday we repeatedly make decisions, which are sometimes logical, and sometimes uncertain. In some fields such as marketing, on the other hand, modeling individuals’ decision process is indispensable for the prediction of the trend. Logical rules however give us no way to represent uncertainties in the decision process, so new tool is required. PRISM (PRogramming In Statistical Modeling) [5] is a programming language for symbolic-statistical modeling, whose theoretical basis is distributional semantics and the learning algorithm for BS programs [4]. In this paper, we conducted an experiment on a car-evaluation database in UCI ML repository [3], in which attributes of various cars are evaluated by a human. The database was originally developed for the demonstration of DEX [1], whose domain has a hierarchy (Figure 1) on ten multi-valued variables: six attributes of a car, three intermediate concepts and the final evaluation, but intermediate concepts are unobservable. Our purpose is to model the decision process and obtain, from estimated parameters, a probabilistic relation between the evaluation and the intermediate concepts