Robust local experiments in sequential process improvement

Abstract

The number of runs in local experiments used in the online improvement of a process need to be small and yet be able to consider as many process parameters as possible. Hence, it would be desirable to perform two-level fractional factorial experiments that involve more factors than just two or three factors at a time recommended in evolutionary operation. In this paper, a case is presented for a careful selection of experimental points that, despite the experiments' being two-level ones, are robust (in their inferences) to nonlinearities in the true response. In order to reduce the chances of obtaining incorrect directions of improvement from fractional factorial local experiments, we incorporate robustness considerations to develop a minimax design procedure for the experiments. The minimax design reduces to the optimal design (under linearity assumptions), if a certain 'nonlinearity-to-noise ratio' is below a threshold. >