Bayesian support vector regression with automatic relevance determination kernel for modeling of antenna input characteristics

摘要

The modeling of microwave antennas and devices typicallyrequires that non-linear input-output mappings be determined between aset of variable parameters (such as geometry dimensions and frequency),and the corresponding scattering parameter(s). Support vector regression(SVR) employing an isotropic Gaussian kernel has been widely usedfor such tasks; this kernel has one tunable hyperparameter that can beoptimized (along with the penalty constant ) using a standard procedurethat involves a parameter grid search combined with cross-validation.The isotropic kernel however suffers from limited expressiveness, andmight provide inadequate predictive accuracy for nonlinear mappingsthat involve multiple tunable input variables. The present study showsthat Bayesian support vector regression using the inherently more flexibleGaussian kernel with automatic relevance determination (ARD) iseminently suitable for highly non-linear modeling tasks, such as the inputreflection coefficient magnitude of broadband and ultrawidebandantennas. The Bayesian framework enables efficient training of the multiplekernel ARD hyperparameters—a task that would be computationallyinfeasible for the grid search/cross-validation approach of standard SVR.