A comparison of ANFIS, MLP and SVM in identification of chemical processes

This paper presents a comparison of Adaptive Neuro Fuzzy Inference Systems (ANFIS),Multilayer Perceptron (MLP) and Support Vector Machines (SVMs) in identification of a chemical process displaying a rich set of dynamical responses under different operating conditions. The methods considered are selected carefully as they are the foremost approaches exploiting the linguistic representations in ANFIS, connectionist representations in MLP and machine learning based on structural risk minimization in SVM. The comparison metrics are the computational complexity measured by the propagation delay, realization performance and design simplicity. It is seen that SVM algorithm performs better in terms of providing an accurate fit to the desired dynamics but a very close performance result can also be obtained with ANFIS with significantly lower computational cost. Performance with MLP is comparably lower that the other two algorithms yet MLP structure has the lowest computational complexity.