Optimizing Direct Learning Neural Network Digital Predistortion Through the Lottery Ticket Hypothesis Agent

This work presents a neural network digital pre distortion optimizer combined with memory backpropagation direct learning architecture. The optimization is done by iteratively pruning low valued weights. The reduced architectures are compared and evaluated on a wideband CMOS power amplifier using a 20 MHz 802.11ac WiFi and 160 MHz QAM64 signal at center frequency of 1 GHz with average powers 10 dBm and 13.5 dBm, respectively. The best optimized models achieve a correction of 13.1 and 11.5 dB to the error vector magnitude from the value of -24.9 and -25.9 dB to the value of -38 and -37.4 dB respectively. The pruned model at 160 MHz signal outperforms the generic model by 0.2 dB EVM and reduces the complexity by 20%, while the optimized model for the 20 MHz signal degrades the generic model performance by 0.1 dB EVM but reduces complexity by 50%.