Surrogate Classification based on Accuracy and Complexity

Abstract

Abstract The prevalence of various machine-learning modeling techniques and numerous possible model configurations generates a long list of unique surrogate forms. Exhaustive enumeration and search for the best surrogate form from a large pool of candidate forms is a non-trivial task. In this work, we aim to assess similarities in modeling capabilities among many different surrogate forms. We examine modeling capabilities for noisy and non-noisy data based on two different surrogate performance metrics. We use a similarity metric to identify similar pairs of surrogate forms, and then group mutually similar forms into distinct families. The similarities among various forms vary depending on the quality of data set and choice of performance metric. This work enables us to exploit families of similar forms to create a reduced search set of contrasting surrogate forms, and facilitate surrogate form selection.