Adaptive Transfer Strategy for Digital Predistortion With Varying Transmission Configurations

Power amplifier (PA) behaviors are highly correlated under different transmission configurations. Utilizing transfer learning to apply effective information from an already trained digital predistortion (DPD) model under a certain configuration to the current configuration can greatly reduce update resources and shorten the transition period. Fine-tuning is an effective method to implement transfer learning in neural networks (NNs). For the trade-off between performance and update resources, this paper proposes an adaptive fine-tuning strategy (ATS) for DPD networks in time-varying transmission configuration scenarios. ATS can determine the location of the fine-tuning operation according to the transmission configuration to obtain better gains with limited resources. Owing to the high fitting efficiency and good interpretability of Kolmogorov-Arnold network (KAN), the block-oriented KAN (BO-KAN) model is also proposed in this paper. Unlike NNs based on multilayer perceptron (MLP), BO-KAN has the ability to transfer information between networks of different capacities. Therefore, the BO-KAN model with a larger capacity can be used for fast updating in complex configurations without training from scratch. Experimental results demonstrate that the BO-KAN with ATS achieves excellent dynamic linearization performance, with low running complexity and acceptable update complexity.