|
Abstract:
|
Bi-Directional Reflectance Distribution Functions are used in many fields including computer animation modeling, military defense (radar, lidar, etc.), and others. This paper explores a variety of approaches to modeling BRDFs using different evolutionary computing (EC) techniques. We concentrate on genetic programming (GP) and in hybrid GP approaches, obtaining very close correspondence between models and data. The problem of obtaining parameters that make particular BRDF models fit to laboratory-measured reflectance data is a classic symbolic regression problem. The goal of this approach is to discover the equations that model laboratory-measured data according to several criteria of fitness. These criteria involve closeness of fit, simplicity or complexity of the model (parsimony), form of the result, and speed of discovery. As expected, free form, unconstrained GP gave the best results in terms of minimizing measurement errors. However, it also yielded the most complex model forms. Certain constrained approaches proved to be far superior in terms of speed of discovery. Furthermore, application of mild parsimony pressure resulted in not only simpler expressions, but also improved results by yielding small differences between the models and the corresponding laboratory measurements.
|
