Coper: A Methodology for Learning Invariant Functional Descriptions

Abstract

Functional descriptions constitute a significant class of goals for machine learning systems. Physical laws are one example of this kind of description. In this paper the COPER methodology is described, which allows discovery of functional descriptions from incomplete observational data. COPER eliminates irrelevant arguments, generates additional relevant argument descriptors (if some are missing), and generates a functional formula. The important feature of this methodology is that it allows testing of relevance for some of the attribute descriptors without varying their values throughout the training events. To do this, it utilizes the property of invariance of meaningful functional descriptions. One of the examples of COPER’s rediscoveries is Bernoulli’s law.