Artificial neural-network-assisted stochastic process optimization strategies

Abstract

This article presents two hybrid robust process optimization approaches integrating artificial neural networks (ANN) and stochastic optimization formalism-genetic algorithms (Gas) and simultaneous perturbation stochastic approximation (SPSA). An ANN-based process model was developed solely from process input-output data and then its input space comprising design and operating variables was optimized by employing either the GA or the SPSA methodology. These methods possess certain advantages over widely used deterministic gradient-based techniques. The efficacy of ANN-GA and ANN-SPSA formalisms in the presence of noise-free as well as noisy process data was demonstrated for a representative system involving a nonisothermal CSTR. The case study considered a nontrivial optimization objective, which, in addition to the conventional parameter design, also addresses the issue of optimal tolerance design. Comparison of the results with those from a robust deterministic modeling/optimization strategy suggests that the hybrid methodologies can be gainfully employed for process optimization.