Design of equiripple FIR filters using a feedback neural network

The weighted least squares design of FIR filters is implemented in terms of a feedback neural network. The proposed neural network is shown to converge to the global minimum in each iteration for the current weighting function, and as the weighting function is adjusted from iteration to iteration, an equiripple design is achieved. The approach is applicable to FIR filters with piecewise-constant amplitude responses, as well as to digital differentiators and Hilbert transformers. The proposed configuration is amenable to analog very-large-scale integration and can, therefore, be used in real-time signal processing.