New Preprint Available. Neural Networks are taking over DPD!
I have a new preprint available for my submission to the 2019 IEEE International Workshop on Signal Processing Systems in Nanjing, China. The paper is titled “Design and Implementation of a Neural Network Based Predistorter for Enhanced Mobile Broadband” and is available here.
In this paper, I use a neural network (NN) to implement digital predistortion (DPD) to correct for power amplifier (PA) nonlinearities. The main contributions are:
- A novel training method where we learn the NN DPD by first modeling the PA with a NN and backpropagating through the PA NN model to update the DPD NN weights.
- Avoiding the commonly used indirect learning architecture (ILA) by using the above method. ILA learns a postinverter and uses this as the DPD, but this is known to converge to a biased solution.
- Reducing complexity of the predistorter when comparing to a memory polynomial
- Testing on RFWebLab
- FPGA implementation of NN DPD and memory polynomials; showing that the NN has reduced implementation complexity.
Machine learning has been popular for classification tasks such as image recognition. Now many wireless researchers are exploring machine learning for tackling a variety of problems such as spectrum sharing.