A Two-Phase Maxi-Min Algorithm for Forward-Inverse Experiment Design

In customer-driven design of systems or products, one has performance targets in mind and would like to identify system design parameters that yield the target performance vector. Since most simulation models predict performance given design parameter values, this identification must be done iteratively through an optimization search procedure. In some cases it would be preferable to find design parameter values directly via an explicit inverse model. Regression and other forms of approximation ´metamodels´ provide estimates of simulation model outputs as a function of design parameters. It is possible to design fitting experiments (DOE´s) that allow simultaneous fitting of both forward and inverse metamodels. This paper discusses the potential for this strategy and shows a simple two-phase DOE strategy using a maxi-min measure of DOE quality