Benchmarking of Learning Architectures for Digital Predistortion

Indirect and direct learning architectures are the two main parameter identification approaches for digital predistortion systems. While the indirect scheme is less complex, its inherent shortcomings are avoided by the direct learning approach. Trying to answer the question whether this advantage of the direct approach can be exploited in terms of measurable linearization-performance improvement in a predistortion platform for advanced power amplifier structures, we present a performance comparison based on laboratory results for wideband high-power Doherty amplifiers. Rather than using single-shot least-squares estimates, each architecture is combined with an adaptive parameter-update scheme to reach the desired performance range and allow for a fair comparison. In conclusion, although the direct learning approach may excel in peak performance, the indirect learning approach achieves virtually the same average performance over linearization runs and has a clear advantage in terms of robustness.