Nonlinear experimental design using Bayesian regularized neural networks

Novel criteria for designing experiments for nonlinear processes are presented. These criteria improve on a previous methodology in that they can be used to suggest a batch of new experiments to perform (as opposed to a single new experiment) and are also optimized for discovering improved optima of the system response. This is accomplished by using information theoretic criterion, which also heuristically penalize experiments that are likely to result in low (nonoptimal) results. While the methods may be applied to any type of nonlinear‐nonparametric model (radial basis functions and generalized linear regression), they are here exclusively considered in conjunction with Bayesian regularized feedforward neural networks. A focus on the application of rapid process development, and how to use repeated experiments to optimize the training procedures of Bayesian regularized neural networks is shown. The presented methods are applied to three case studies. The first two case studies involve simulations of one and two‐dimensional (2‐D) nonlinear regression problems. The third case study involves real historical data from bench‐scale fermentations generated in our laboratory. It is shown that using the presented criteria to design new experiments can greatly increase a feedforward neural network's ability to predict global optima. © 2007 American Institute of Chemical Engineers AIChE J, 2007