Real-coded genetic algorithms and nonlinear parameter identification

Abstract Macroscopic models are useful for example in process control and optimization. They are based on the mass balances describing the flow conditions and the assumed reaction scheme. The development of a process model typically has two steps: model structure selection and parameter identification. The structure selection step is not discussed in this report. In the parameter identification step, the squared error criterion is typically minimized. It can be done with conventional methods such as gradient methods but genetic algorithms is used in this study instead. That is because genetic algorithms are very likely to find the global minimum as the conventional methods may stuck in local minimums. Real-coded genetic algorithms are used in this study. Thus the first part of this report reviews the crossover and mutation operators used for regulating the development of population. A process simulator of a bioprocess is used to generate data for parameter identification. The bioprocess model is known to be nonlinear having two separate operating points. Thus the process model is identified in two operating conditions. The mean squared error criterion is used to determine the validity of possible solutions to optimization problem. Parameter identification with genetic algorithms performed well giving good results. The optimizations are repeated 500 times to guarantee the validity of the results. The results are further validated by examining the mean and the standard deviation of the prediction error as well as the mean squared error criterion. Also correlation coefficients are calculated and the histograms of the parameter values in all 500 optimizations are studied