A recurrent neural network for online design of robust optimal filters

Abstract

A recurrent neural network is developed for robust optimal filter design. The purpose is to fill the gap between the real-time computation requirement in practice and the computational complexity of the filter design in the case that the statistical properties of noise are unknown. First, an H/sub /spl infin// requirement and an L/sub 2/ requirement of the filter design problem are formulated as a group of linear matrix inequalities. On this basis, an optimization problem is introduced to solve the robust optimal filter design problem. Then, a recurrent neural network is deliberately developed for solving the optimization problem in real time. The effectiveness and efficiency of the recurrent neural network is shown by use of theoretical and simulation results.