Early Stopping Criterion for Recursive Least Squares Training of Behavioural Models

Abstract

As the physical makeup of cellular base-stations evolve into systems with multiple parallel transmission paths the effort involved in modelling these complex systems increases considerably. One task in particular which contributes to signal distortion on each signal path, is the power amplifier. In power amplifier modelling, Recursive Least Squares has been used in the past to train Volterra models with memory terms, however instability can occur when training the model weights. This manuscript provides a computationally efficient technique to detect the onset of instability and subsequently to inform the decision when to stop adaptive training of dynamic nonlinear behavioural models and avoid the onset of instability. This technique is experimentally validated using four different signal modulation schemes.