Early Stopping Criterion for Recursive Least Squares Training of Behavioural Models

The necessity of the rapid evolution of wireless communications, with continuously increasing demands for higher data rates and capacity Zheng (Big datadriven optimization for mobile networks toward 5g 30:44–51, 2016), is constantly augmenting the complexity of radio frequency (RF) transceiver architecture. A significant component in the configuration of such complex radio transceivers is the power amplifier(PA). Multiple distributed PAs are now common in proposed RF architectures. PAs exhibit non linear behaviour, causing signal distortion in transmission. Behavioural models offer a concise representation of a PAs characteristic performance which is extremely useful in simulating performance of multiple nonlinear power amplifiers. A considerable drawback with using the Recursive Least Squares (RLS) technique is that the instability of the coefficients during the training of the model. This manuscript provides a computationally efficient technique to detect the onset of instability during adaptive RLS training and subsequently to inform the decision to cease training of dynamic memory polynomial based behavioural models, to avoid the onset of instability. The proposed technique does not require modification of the RLS algorithm, merely an observation of the pre-exsisting autocorrelation function based update. This technique is experimentally validated using four different signal modulation schemes, LTE OFDM, 5G-NR, DVBS2X and WCDMA.