NEURO-FUZZY SYSTEMS FOR MODELLING AND CONTROL APPLICATIONS

Abstract

The M. Tech Dissertation work entails the brief overview of Adaptive Neuro-Fuzzy Inference Systems (ANFIS) Architecture and its conventional Hybrid Learning Algorithm (HLA). Hybrid Learning Algorithm uses two passes (forward and backward pass), which is a combination of Least Square Estimate (LSE) and back propagation based on gradient descent. As it uses gradient based method, it has certain drawbacks such as the calculation of gradient is possible with differentiable membership functions only, over fitting, calculation complexity and an iterative method hence time consuming. To overcome disadvantages of Hybrid Learning Algorithm, to improve the flexibility of ANFIS Architecture, and to improve learning speed of network. The new learning mechanism Extreme-A NFIS is proposed. The proposed algorithm is designed on the basis of the concept of Extreme Learning Machine (ELM) which is fastest learning mechanism designed for single-hidden layer feed-forward neural networks (SLFNs) by G.-B. Huang. In spite of the advantages, ELM suffers from the inherent randomness of the results and ANFIS has strong computational complexity restrictions because of hybrid learning algorithm. The Extreme-ANFIS reduces the computation complexity of the ANFIS by eliminating the hybrid learning algorithm and avoids the randomness of the ELM networks by incorporating explicit knowledge representation using fuzzy membership functions. The thesis also includes the rigorous analysis of applications like modelling, classifi- - cation and control. The comparative analysis with other method are used to derive some conclusive features of proposed mechanism. The design of extreme-ANFIS Learning algorithm and its applications are developed in MATLAB environment