Adaptive filter design : An information theoretic learning approach

Abstract

Adaptive _ltering has gained wide popularity in recent times in non stationary signal processing environment. Supervised adaptive _ltering requires the use of a reference signal and an adaptive algorithm. The robustness of adaptive algorithms are put to real test while operating in real time environments contaminated with practical alpha stable noise. The conventional algorithms tend to loose stability easily, resulting in improper or diverging learning results. The use of _nite order statistics is cited as the major reason for this behaviour. The information theory, a popular _eld in communication engineering is gaining wide acceptance in many conventional signal processing problems. This thesis tries to exploit the merits of correntropy, which is related to correlation and entropy, and use the same in adaptive _ltering by analysing some practical systems namely noise cancellers, generalised sidelobe cancellers and active noise control\ systems. Practical implementation on a standard DSP processor has been done to see the behaviour of noise canceller in real time. Rigorous analysis has been carried out to _nd out the merits of such systems supplemented by information theoretic learning against conventional second order statistics based learning.