Designing an equalizer structure using gradient descent algorithms

Abstract

In recent years, a growing field of research in “Adaptive Systems” has resulted in a variety of adaptive automatons whose characteristics in limited ways resemble certain behaviors of living systems and biological adaptive processes. The essential and principal property of the adaptive systems is its time-varying, self-adjusting performance by using a process called “learning” from its environment. A channel equalizer is a very good example of an adaptive system, which has been considered in this work to assess its performance with reference to various novel learning algorithms developed. The two main threats for the digital communication systems are Inter-symbol Interference (ISI) and the presence of noise in the channels which are both time varying. So, for rapidly varying channel characteristics, the equalizer too need to be adaptive. In order to combat with such problems various adaptive equalizers have been proposed. Particularly, when the decision boundary is highly nonlinear, the classical equalizers (so called linear ones) do not perform satisfactorily. The use of Artificial Neural Networks (ANNs) provides the required nonlinear decision boundary. The Back Propagation (BP) algorithm revolutionized the use of ANNs in diverse fields of science and engineering. The main problem with this algorithm is its slow rate of convergence. But the high speed digital communication systems, in the presence of rapidly fading channels, demand for faster training. To overcome this problem a faster method of training the neural network using RLS algorithm is proposed in this thesis work. But both the BP and RLS based BP algorithms belong to the family of Gradient-based algorithms, which have the inherent problem of getting trapped in local minima. Since obtaining a global solution is the main criterion for any adaptive system, an efficient search technique is highly desirable. Tabu Search serves this purpose.