The use of boundary conditions for inductive models

Abstract

There is a large amount of interest in creating models from data using a variety of machine learning methods. Most of these approaches require a good distribution of observed values to produce reliable models. The use of background knowledge to augment the observed values has also been explored as a method to supplement the original feature set of training data. This paper argues that there is an additional set of data that can be created for many types of problems, based on the concept of boundary conditions. This boundary data incorporates an understanding of the modeled system behaviour under certain extreme values and therefore reduces the degrees of freedom within the inferred model. This paper argues that by using this information when training an inductive model a more robust generalization of the data can be achieved under some circumstances.