Constructing models of hidden structure

Abstract

The search for structure hidden inside of visible objects has been a universal research task in Physics, Chemistry, Biology, Philosophy, and other disciplines. Discovering hidden structure is an especially challenging type of constructive induction. Not only must all the elements of hidden structure be postulated by the discoverer, but they can only be verified by indirect evidence, available at the level of observable objects. In this paper we describe a framework for the automation of hidden structure discovery. We define what is meant by hidden structure and we present a number of operators that can build models of hidden structure step by step. Our models of hidden structure consist of hidden objects of several types, admissible combinations of hidden objects, the attributes of hidden objects and their combinations, a mapping between the hidden and the observed structure, and reactions described in terms of hidden objects. We analyze the discovery of atoms, genes, and quarks to demonstrate the generality of our operators. We demonstrate how domain knowledge on the visible level is useful in operator instantiation. We discuss efficient control structures, and we define the criteria for model evaluation. Because hidden structure cannot be verified by direct observations, a successful model must pass two stages of evaluation. First, the observational consequences must be confirmed, and second, the model must be unique in its simplicity class.