الفهرس 
introductionOnly 14 pages are availabe for public view
 |  | from | 16 |  |  |
| | | from | 16 | | |
Abstract
This thesis is concerned with analog Very Large Scale Integration (VLSI) implementation of . Adaptive Neuro-Fuzzy Inference System (ANFIS). ANFIS has shown superior performance over other computational paradigms like Back-Propagation Neural Networks, Cascade Correlation Neural Networks and Linear Predictive Methods. By superior. performance we mean better generalization capability with fewer adjustable parameters and real time adaptation. ANFIS is suitable for many applications such as nonlinear function modeling, system identification, adaptive control and time series prediction. The thesis discusses fuzzy inference systems and on-chip learning neural networks using st6chastic preservative. techniques which are more VLSI friendly than standard learning . techniques such as back-propagation. The discussion includes reporting the results of circuit level simulation of designed structural. units which are assembled together to build a fuzzy logic controller, an on-chip learning neural network and finally the combined neuro-fuzzy inference system.. These blocks. include voltage to current converters, current to voltage converters, adders, analog multipliers, squaring circuits, square root circuits, analog dividers, minimum current circuits, membership function circuits, level shifters, comparators and analog storage circuits with adaptation mechanism. A subset of these blocks is combined to build a 5-rulewaterlevel fuzzy logic controller that is simulated and the results of simulation are compared with the sarne fuzzy system implemented using MATLAB. Another subset is used to build a neural network to perform XOR task and the results of simulating this ’network are presented with emphasis on the on-chip learning capability. Then the main thesis
objective is presented which is combining the neural and fuzzy systems to build on-chip learning ANFIS for the purpose of predicting the Mackey-Glass chaotic time series. The prediction of future values of this time series is a benchmark problem that has been used and reported by a number of researchers. The obtained results’ show how on-chip learning is very fast compared to software implemented learning algorithms. MA TLAB, running on an advanced platform, learns the prediction task in several minutes while the hardware implemented circuits, as the circuit simulation reveals, can learn the job in fraction of a second with non-noticeable lesser accuracy due to differences in the membership functions and algorithms used in both techniques. The thesis demonstrates the potential of neuro-fuzzy hardware with on-chip learning capability in the field of adaptive signal processing and control.