Two-stage computing budget allocation approach for response surface method

Abstract

Response Surface Methodology (RSM) is one of the main statistical approaches to search for the input combination that optimizes the simulation output. In the early stages of RSM, the steepest ascent method, which is scale-dependent, is used. Recently, a novel technique is proposed to identify the next design point with a scale-independent search direction and a possible step size. In this thesis, we improve this technique by considering the more realistic case where there are computing budget constraints, and look into the important question of experimental design. We propose a two-stage computing budget allocation approach, and do several numerical experiments to compare it with the traditional factorial designs which allocate budget equally to each design point. The results show that our technique outperforms the traditional equal allocation when the system noise is large.