Genetic Programming for Symbolic Regression of Chemical Process Systems

The novel evolutionary artificial intelligence formalism namely, genetic programming (GP) a branch of genetic algorithms is utilized to develop mathematical models based on input-output data, instead of conventional regression and neural network modeling techniques which are commonly used for this purpose. This paper summarizes the available MATLAB toolboxes and their features. Glucose to gluconic acid batch bioprocess has been modeled using both GPLAB and hybrid approach of GP and Orthogonal Least Square method (GP OLS). GP OLS which is capable of pruning of trees has generated parsimonious expressions simpler to GPLAB, with high fitness values and low mean square error which is an indicative of the good prediction accuracy. The capability of GP OLS to generate non-linear input-output dynamic systems has been tested using an example of fed-batch bioreactor. The simulation and GP model prediction results indicate GP OLS is an efficient and fast method for predicting the order and structure for non-linear input and output model. However, if an accurate process model were available, then many of the benefits of improved process operability would be achievable. The current trend within the process industries is to use data based modeling techniques to develop accurate, cost-effective input-output process descriptions (3). The popular techniques may be divided into two categories. The first are based on the use of various statistical techniques and regression analysis, while the second involves the use of artificial neural networks.