the use of millimeter wave (mmWave) spectrum, it addresses the issue of need for more bandwidth and high data rate because of its nature. For example, when we consider mmWave precoding, the use of DL for predicting the precoding matrices has shown to reduce the channel training overhead when compared traditional singular value decomposition methods [9]. The computational complexity of the problem is taken care of by using DL, because of its ability of learning from data. The applications of wireless communications involving real-time scenarios such as data monitoring and video streaming, which play a major part of modern day, have improved a lot today because of the growing usage and new research being done in DL. The classification of digital and analog modulating signals is done with some key features in [25]. DL/ML has wide range of applications in the wireless communication systems as can be seen in Figure 5.1. The some of the most relevant and in scope of this study are discussed in detail. DL as discussed previously is a subbranch of ML and is considered in our application because of its capability to handle much larger data sets, creating a more complex neural network (NN). The major difference between them is that ML is used for small signal data, whereas DL is used for high dimensional data and has a good performance in real-time scenarios. Some popular algorithms in DL are convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTMs), stacked auto-encoders, deep Boltzmann machine (DBM), and deep belief networks (DBNs) [26]. Every algorithm is uniquely designed and has many wide range of applications in wireless communication. A few of them are discussed in the subsequent subsections. These wireless applications are chosen based on their popularity and suitability for ML/DL algorithms.
Figure 5.1 ML application for communications (re-generated from [24]).
5.2.1 Automatic Modulation Classification
AMC is the process of classifying the modulation scheme of a signal and is a core technique of non-cooperative communication. It is an intermediary step between signal reception and signal demodulation. It is essential for rapid response and signal identification in dense electromagnetic spectrums in the presence of channel noises and multipath effects in military, cognitive radio, and 5G network applications. The two key steps in the design of a modulation classifier are signal pre-processing and selection of an effective classification algorithm. The pre-processing section might include (but are not limited to) reduction of signal noise, estimation of carrier frequency and signal power and extraction of essential signal information (as per the requirements of the classification algorithm) [27]. As for the second part, modulation classification algorithms can be either likelihood based, or feature based. Likelihood-based classifiers essentially work by comparing the likelihood ratio of the received signal against a predetermined threshold and have the advantage of minimizing the probability of a false classification but are more complex, require higher computational power, and are more difficult to implement in hardware. Feature-based classification has two main subsystems 1: feature extraction (which can be considered a pre-processing step) [28] and 2: classification [29, 30]. This method makes use of certain specific features of the received signal and classifies based on these feature values. Although they have less than optimal performance, feature-based classifiers greatly reduce computational complexity and can be acceptable with proper design.
A wide amount of work has been done in feature-based AMC and the application of DL algorithms directly on the received signal, thereby eliminating the feature extraction step and further reducing computational complexity. The application of CNN in AMC has shown promising accuracy which can ensure acceptable performance with much lower cost of computation. The next step would be to find effective hardware implementation of DL-based AMC classifiers [31].
5.2.2 Resource Allocation (RA)
The RA problem in wireless communication systems is considered as one of the most challenging tasks. The RA problem is formulated as an optimization problem and usually solved online with available information [32]. It is difficult to obtain a real-time optimal solution for most RA problems due to their nonconvex nature. To solve these problems, Lagrangian and greedy methods are employed which results in performance degradation [33]. The nonlinear programming (NLP) methods were used to solve the RA problem, due to their cubic complexity, the implementation of these methods were also targeted on graphics processing units (GPUs) for faster processing [34]. Hence, the traditional algorithms for RA are facing great challenges in achieving the QoS requirement of the users in scarce wireless scenarios. RA has a great ability to provide a guaranteed user’s QoS by optimizing the available facilities to minimize operational cost and maximize the operator’s revenue. Therefore, the efficient RA is always a trending topic for future wireless communication networks.
In recent years, there has been a drastic increase in internet traffic and expected to grow in future wireless systems [2, 35]. This traffic growth contributed by the various applications such as wide variety of user equipment (UE), smartphones, automatic vehicles, and IoT sensors. Due to this enormous growth in internet traffic, radio RA in future wireless networks (5G and beyond) is becoming more challenging. Therefore, RA resurfaced as a trending topic in the wireless communication area [36]. DL methods have a great potential to efficiently optimize the radio resource in future wireless systems. Recently, Zhou et al. [37] proposed a DL-based radio RA in ultra-dense 5G networks. In [37], authors have proposed LSTM method for RA problem in 5G scenario and achieved low packet loss along with high throughput. Wang et al. [38] and Zhang et al. [39] presented ML-based RA problems assisted with cloud computing. DL has shown great potential and provided a break-through in a variety of research areas [21].
5.2.3 Channel Estimation/Signal Detection
The application of a neural network (NN) for channel estimation is influenced by the channels which are challenging to describe. This problem may ensue from a provision that inhibit the possession of CSI at the receiver (CSIR) or an unavailability of a well-known channel models. For instance, in MIMO systems with low-resolution analog-to-digital converters (ADCs), consistent CSIR cannot be achieved due to the abrasive quantization instituted by the ADC [22]. In molecular communication systems, the fundamental channel models are indefinite [40]. Therefore, both systems offer themselves to NN-based detectors. Jeon et al. [40] draw a contrast in speech recognition and found that this a domain in which DL algorithms have done extremely good. Speech recognition and digital communication both begin with a signal which is generally sent over a channel to some receiver. This channel can be a wireless, acoustic, or chemical and a receiver can be a microphone, cell phone, or chemical sensor. The receiver tends to detect the original transmitted signal. This evaluation emphasizes the ability of DL algorithms in signal detection over undetermined channels.
In wireless detection, we initially estimate the parameters of a channel over which the signal is being transmitted. These estimates of the CSI are needed for detection at the receiver. Conventional algorithms to estimate CSI, such as minimum mean square error (MMSE) or maximum a posteriori probability (MAP) estimation, necessitate an analytical model of the channel, but the blend of channel distortion and hardware imperfections can be challenging to model systematically. Authors in [41] employed a DL-based method to estimate the carrier frequency offset (CFO) and timing estimates to empower detection of single carrier phase-shift keyed signals.
5.2.4 Millimeter Wave
The mmWave communication serves as a promising solution to improve the spectral efficiency and the problem of shortage of frequency resource [42]. Many applications are related to mmWave communication such as beamforming, precoding, and channel estimation. Power allocation are used for mmWave with non-orthogonal multiple access (NOMA)–related applications. The blockages when mmWaves are being used are
