Neural networks for optimization of nonquadratic cost functions with application to adaptive signal processing

In recent papers [4, 5], a new neural adaptive filtering structure has been proposed, based on a Least Squares (LS) performance function of errors. In this work, the structure in [4, 5] is generalized and a neural adaptive FIR filter is designed whose performance function is expressed in the general non-LS form. The proposed neural filter is shown to compute in real time (i.e. within each sampling interval) the optimal set of the programmable weights for general non-LS cost functions. As a consequence it features excellent tracking capabilities and is effective for on-line applications where fast adaption speed is required. It is also shown that for some common non-LS cost functions, the neural structures here proposed can be implemented on relatively simple electronic circuits that can be fully integrated in MOS VLSI technology. © 1992 IEEE.