Faster convergence and improved performance in least-squares training of neural networks for active sound cancellation

Abstract

This paper introduces new recursive least-squares algorithms with faster convergence and improved steady-state performance for the training of multilayer feedforward neural networks, used in a two neural networks structure for multichannel non-linear active sound cancellation. Non-linearity in active sound cancellation systems is mostly found in actuators. The paper introduces the main concepts required for the development of the algorithms, discusses why it is expected that they will outperform previously published steepest descent and recursive least-squares algorithms, and shows the improved convergence produced by the new algorithms with simulations of non-linear active sound cancellation.