Guiying Zhang, Food Science & Technology, University of California, One Shields Avenue, Davis, CA 95616, Matthew M. Olsen, Department of Chemical Engineering and Materials Science, University of California, One Shields Avenue, Davis, CA 95616, and David E. Block, Department of Chemical Engineering and Materials Science and Department of Viticulture and Enology, University of California Davis, One Shields Ave, Davis, CA 95616.
To identify the optimum of a multi-dimensional system, a novel nonlinear experimental design (DOE) scheme was computationally evaluated. The method, a hybrid neural network-genetic algorithm (NN-GA) coupled with a fuzzy clustering technique, was developed to accumulate the information from former experiments and suggest optimal operating conditions for future experimentation. Due to the uncertainty associated with complex systems, several surfaces with various degrees of complexity and dimensionality were used as case studies for evaluating the new methodology to gain insight into the challenges that may be presented by arbitrary experimental optimization problems likely to exist in practice. The long-term objective of this work is to find better process operating conditions for any difficult optimization problem in fewer experiments. The simulation results demonstrate that a practical number of experiments in a round of experimentation and a reasonable degree of convergence in the genetic algorithm improve the results of the new scheme proposed, but higher complexity, dimensionality, and noise levels decrease its performance. However, even at higher levels of complexity, dimensionality, and noise, this new approach finds significantly better solutions than those found with traditional statistical experimental design methods, with similar or only slightly more experiments.