Process design optimisation using embedded hybrid visualisation and data analysis techniques within a genetic algorithm optimisation framework

Process optimisation is a difficult task due to the non-linear, non-convex and often discontinuous nature of the mathematical models used. Although significant advances in deterministic methods have been made, stochastic procedures, specifically genetic algorithms, provide an attractive technology for solving these optimisation problems. However, genetic algorithms are not naturally suited to highly constrained problems. We propose a targeted genetic algorithm for process optimisation which is suitable for highly constrained problems. The genetic operators, crossover and mutation, are defined based on information gained about the feasible region and the behaviour of the objective function through the use of a data analysis procedure. The data analysis is based on a visual representation, the parallel co-ordinate system. A pattern matching algorithm, the Scan Circle Algorithm [K. Wang, A. Salhi, E.S. Fraga, Cluster identification using a parallel co-ordinate system for knowledge discovery and nonlinear optimization, in: J. Grievink, J. van Schijndel (Eds.), Proceedings of the 12th European Symposium on Computer-Aided Process Engineering, Computer-Aided Chemical Engineering, vol. 10, Elsevier, Amsterdam, 2002, pp. 1003–1008], is extended through the use of Learning Vector Quantization [T. Kohonen, Self-Organizing Maps, Springer-Verlag, Heidelberg, 1995] to identify, automatically, key features of the objective function and the search space. These features are used to target the genetic operators. Results from the application of the new targeted genetic algorithm to an oil stabilisation problem are presented, demonstrating the effective, efficient and robust nature of the implementation. The use of visualisation as the core of the data analysis step also provides a useful tool for explaining the results obtained by the optimisation procedure.