The generalization performances of the Back Propagation Multi-Layer Perceptron (BPMLP) and the Radial Basis Function (RBF) neural networks were compared together by resorting to several sets of experimental data collected from a pilot scale packed absorption column. The experimental data were obtained from an 11cm diameter packed tower filled with 1.8 meter ¼ inch ceramic Rashig rings. The column was used for separation of carbon dioxide from air using various concentrations and flow rates of Di-Ethanol Amine (DEA) and Methyl Di-Ethanol Amine (MDEA) solutions. Two efficient algorithms were employed for optimal training of both neural networks. The Leave One Out Cross Validation (LOOCV) criterion was employed to compute the optimum level of regularization for RBF networks. An in-house procedure was also exploited to predict the optimal widths of isotropic Gaussian basis functions for these networks. Another in-house algorithm was used to train the MLP networks more rapidly and efficiently than the conventional procedures. The simulation results indicated that the RBF networks can perform more adequately than the MLP networks. Because, the RBF networks enjoy a solid theoretical background which enables them to successfully filter out the noise and provide more reliable generalization performances.

Keywords