A hybrid differential evolution self-organizing-map algorithm for optimization of expensive black-box functions

Shreyas Subramanian

Abstract: DIRECT application of optimization algorithms is not practical for expensive black-box functions. Existing meth-ods often require more time or function evaluations than what can be afforded.[1] Surrogate models have been widely used to overcome this difficulty.[1, 2, 3, 4, 5]. Surrogates are fast approximations constructed using data from these expensive black-box functions, and make optimization studies feasible. Typical issues that arise in the surrogate based analysis and optimization (SBAO) of computationally expensive models are exposed in Queipo et al.[6] The issues addressed include selection of the regularization criteria for constructing surrogates, the need for successive approximations, poor convergence rates and sub-optimal values obtained at convergence.. Furthermore, non-interpolating surfaces are unreliable since the shape of the function may not be sufficiently captured.