A Generalized Digital Predistortion Model Based on Artificial Neural Networks

In this paper, the CPWL-based DPD model in is extended to a DPD model based on artificial neural networks, where the mapping from R-n to C-m in CPWL functions is implemented by ANNs. The linearization performance of ANNs model with different activation functions is also validated with platform tests. The test results show the ANNs-based DPD model with ABS activation function can achieve similar performance as the CPWL-based DPD model. The performance of ANNs model with other activation functions (PWL, ReLU and tanh functions), is 1 similar to 2 dB worse than that of the ABS activation function.