default Underwater acoustic link adaptation based on regression trees


Underwater acoustic link adaptation based on regression trees. Pelekanakis, Konstantinos; Cazzanti, Luca; Fountzoulas, Yannis; Alves, João. CMRE-FR-2018-005. November 2018.

Adaptive Modulation and Coding (AMC) is considered an efficient way to achieve high spectral efficiencies in non-stationary underwater acoustic channels. In this work, we tackle the problem of maximising the bit rate of a single-carrier communications system under a target BER constraint. The modem is equipped with a set of Phase-Shift Keying (PSK) signals of various baud rates and modulation orders. These PSK signals were transmitted during REP15-Atlantic off the coast of Portugal for a period of five days. Receiver processing is based on a channel-estimate-based Decision Feedback Equalizer (DFE). Since analytical formulas for the BER computation are intractable, we turn our attention to machine learning techniques that can predict the BER directly from the data. In particular, we compare the BER prediction error of decision trees, boosted trees, and random forests. Our work deals with selecting the appropriate split of training and test sets and finding the tree parameter values for optimal prediction. We found that boosted trees yielded the best BER prediction performance. Furthermore, all three investigated tree models achieved near-optimal prediction for the same parameter values. The practical significance is that a robust choice of parameters can make the tree model generic enough to maintain performance over long periods. Finally, the efficiency of the employed trees to adapt the physical layer is shown in post-processing.