Home // SIMUL 2016, The Eighth International Conference on Advances in System Simulation // View article

A Comparison of Some Simple and Complex Surrogate Models: Make Everything as Simple as Possible?

Authors:
Wim De Mulder
Bernhard Rengs
Geert Molenberghs
Thomas Fent
Geert Verbeke

Keywords: Gaussian process emulation; Radial basis function networks; Inverse distance weighting; Agent-based models; Cluster analysis.

Abstract:
In this paper we compare three surrogate models on a highly irregular, yet real-world data set. The three methods strongly vary in mathematical sophistication and computational complexity. First, inverse distance weighting, a very intuitive method whose single parameter can be readily determined via basic, albeit nonlinear, optimization. Secondly, radial basis function networks, for which some parameters can be determined via simple matrix algebra, although determination of other parameters require cluster analysis and nonlinear optimization. Thirdly, Gaussian process emulation, a statistical technique having a technical mathematical formulation and where specification of parameter values rely on complex and time-consuming optimization. It comes as a complete surprise that inverse distance weighting performs best on our complex data set. Our work encourages to moderate the extreme optimism about and overly use of very advanced methods. The commonplace a priori assumption that simple, intuitive methods underperform on complex data sets is not always justified, and such methods still have their place in the current era of highly advanced computational and mathematical models.

Pages: 26 to 32

Copyright: Copyright (c) IARIA, 2016

Publication date: August 21, 2016

Published in: conference

ISSN: 2308-4537

ISBN: 978-1-61208-501-2

Location: Rome, Italy

Dates: from August 21, 2016 to August 25, 2016