Ensemble Learning Methods for Full-Duplex Self-Interference Cancellation

Self-interference cancellation (SIC) is a major challenge to be accomplished in full-duplex (FD) communication systems. Due to nonlinear impairments introduced by components of the FD transceiver chain, estimating the self-interference signal is computationally demanding. Recently, machine learning methods have been developed for SIC, and have shown to outperform model-based methods in terms of computational complexity. To further reduce complexity, in this work we investigate the use of ensemble learning methods with regression trees as base learners for SIC, as these methods are particularly suitable for efficient hardware implementation. Further, we propose a novel neural network (NN)-based ensemble learning method. The presented approaches achieve state-of-the-art performance with considerably lower complexity compared to a model-based polynomial approach and a single fully-connected NN.