The main goal of the project/thesis was to model a physical layer of a simple wireless communications system with Autoencoders. During the phase of the research, I was able created a new mathematical model of 4PAM constellation mapping which was subjected non-linear noise rather than usual gaussian noise (AWGN). For higher order constellation, deep learning was used to learn the constellation mapping that would adopt to the non-linear noise introduced.

Deep Learning has been known to optimize real world problems with robust mathematical models. One of such applications could be modeling a physical layer for Wireless Communcations using Neural Networks (NN). This thesis proposes the use of Autoencoder to represent a simple communications system. During transmission of a signal, non-linearity is introduced due to hardware impairments and amplification of signal. Also, self-interference caused from full-duplex communication system adds non-linearity. In these cases, the con- stellation from conventional modulation schemes are not optimal. This thesis will focus on modeling such distortions as non-linear Additive- White Gaussian Noise (AWGN) channel and finding the optimal constellation to overcome this non-linearity using deep learning methods. The quality of the NN is assessed with the help of Bit Error Rate (BER) plots. We see that the Autoencoder is able to match the performance of standard constel- lation in linear noise while it outperforms those constellation in non-linear noise. The results are then verified using analysis on constellation of simple Pulse-amplitude Modulation (PAM).

├── docs
│   └── Bachelors-Thesis-BK.pdf
├── README.md
└── src
    ├── 16QAM
    │   └── 16QAM-AEC.ipynb
    └── 4PAM
        ├── 4PAM-AEC.ipynb
        └── matlab
            └── BFO.m

To cite this work use the following:

@mastersthesis{nlwc,
    author={Khanal B.},
    title={Non-Linearity in Wireless Communications and Deep Learning},
    year= {2020},
    type={Bachelor's thesis},
    school={Jacobs University Bremen}
}