We compare two approaches where empirical models are used to augment computer simulations to facilitate rapid device optimization. We apply response surface model (RSM) methodology and neurofuzzy techniques to the problem of modeling simulations of the average flux density in the air gap of a loudspeaker. Both these techniques have significant advantages over more traditional methods of optimizing computer simulation experiments. We show that these techniques have different advantages and disadvantages depending on the problem being modeled. In particular, the use of domain knowledge is shown to give robust and reliably predictive RSM's. Neurofuzzy techniques are shown to be particularly suited to problems where little is known about the problem.Index Terms-Design of experiments, flux density and computational intelligence, modeling, neurofuzzy models, optimization, response surface models, simulation.

No abstract

No abstract

This paper presents the case against two widespread practices in designing engineering experiments, which are (i) to vary one factor at a time (the OFAT approach) and (ii) to generate experimental points by random selection (the Monte Carlo approach). These approaches do not produce good experimental designs, defined as generating maximum information per run (IPR), and should be replaced by designs that do. These latter designs (i) vary many factors at a time and (ii) use a patterned set of experimental points rather than a random set. An example from circuit design is used to illustrate this approach. The limitations of the random approach are well known amongst statisticians but often not among engineers. © 1998 John Wiley & Sons, Ltd.

No abstract