Machine Learning Approaches for Faster-than-Nyquist (FTN) Signaling Detection
There will be a significant demand on having a fast and reliable wireless communication systems in future. Since bandwidth and bit rate are tightly connected to each other, one approach will be increasing the bandwidth. However, the number of wireless devices are growing exponentially, and we don't have infinite bandwidth to allocate. On the other hand, increasing the bit rate for a given bandwidth, i.e., improving the spectral efficiency (SE), is another promising approach to have a fast and reliable wireless communication systems. Faster-than-Nyquist (FTN) is one of the candidates to improve the SE while this improvement comes at the expense of complexity of removing the introduced inter-symbol interference (ISI). In this thesis, we propose two algorithms to decrease the computational complexity regarding removing the ISI in FTN signaling. In the first main contribution of the thesis, we introduce an equivalent FTN signaling model based on orthonormal basis pulses to transform the non-orthogonal FTN signaling transmission to an orthogonal transmission carrying real-number constellations. Then we propose a deep learning (DL) based algorithm to decrease the computational complexity of the known list sphere decoding (LSD) algorithm. In essence, the LSD is one of the algorithm that can be used for the detection process of the FTN signaling; however, at huge computational complexity. Simulation results show the proposed DL-based LSD reduces computational complexity by orders of magnitude while maintaining close-to-optimal performance. In the second main contribution of the thesis, we view the FTN signaling detection problem as a classification problem, where the received FTN signaling signal viewed as an unlabeled class sample that is an element of a set of all potential classes samples. Assuming receiving $N$ samples, conventional detectors search over an $N$-dimensional space which is computationally expensive especially for large value of $N$. However, we propose a low-complexity classifier (LCC) that performs the classification in $N_p$ dimensional space where $N_p\ll N$. The proposed LCC's ability to balance performance and complexity is demonstrated by simulation results.
Faster-than-Nyquist, Machine learning
Master of Science (M.Sc.)
Electrical and Computer Engineering
Stakhanova, Natalia;Bui, Francis;Bedeer Mohamed, Ebrahim