Global Optimization of Problems with Disconnected Feasible Regions via Surrogate Modeling

This work proposes taking advantage of superEGO’s flexibility to design alternative search strategies that locate multiple feasible regions in a new way, and was successfully applied to two analytical examples and required far fewer function calls than three competing techniques.

Abstract

Approximation methods have successfully solved a variety of global optimization problems. One of the major remaining challenges involves problems where the feasible design space consists of disconnected regions. Traditionally, researchers have attempted to solve such problems by combining global and local searching algorithms such as Simulated Annealing and SQP. While this approach is often successful, it is not ecient in terms of the number of function evaluations required. When the design problem is comprised of expensive functions such as computer simulations, new techniques must be applied. We use an approximation-based global optimization algorithm, superEGO, to solve such problems. We propose taking advantage of superEGO’s flexibility to design alternative search strategies that locate multiple feasible regions in a new way. The technique was successfully applied to two analytical examples and required far fewer function calls than three competing techniques.