Incorporating machine learning algorithms for channel prediction and power optimization scheme in power domain non-orthogonal multiple access system

Abstract

With the fast increase in the publicity of the Internet of Things (IoT) and cloud computing, the requirement for massive connectivity and highly reliable data rates is increasing day by day for communication networks. IoT can establish the connections among many types of smart devices, such as smart sensors, robots, and mobile devices. To satisfy these demands and such massive connectivity, three main services have been presented to the communication networks. These key services consist of massive machine type communication (mMTC) that permits massive connections between IoT terminals, enhanced mobile broadband (eMBB) that delivers a high data rate for mobile devices, and ultra-reliable and low-latency communication (URLLC) that confirms reliability and minimum latency for critical and sensible applications. These services are characterized by their quality of service (QoS), where URLLC has a stringent QoS policy for high reliability and low latency application, eMBB service is categorized by a moderate QoS policy, while mMTC has no precise QoS policy. These types of QoS are usually difficult to realize with the traditional orthogonal multiple access (OMA) due to limited spectrum resources, and delays. To satisfy and enhance these diverse QoS requirements, many potential multiple access schemes have been introduced into communication network. Among them, is the non-orthogonal multiple access (NOMA) scheme that has a achieved a popularity because it can support massive connectivity with limited resources, tolerable transmission delays, and high spectral efficiency. The key feature of NOMA is that multiple user devices can be served from the same radio resource block, such as time, frequency, and code. NOMA scheme applies superposition coding to combine signals related to multiple users at the transmitter side and implements successive interference cancellation procedure to differentiate and recover the signals of multiple devices at the receiver side. There are some challenges related to resource allocation in NOMA system, such as power allocation and channel estimation. Machine learning has obtained publicity over the past several years, and many machine learning models and algorithms have been industrialized. Deep learning is a subset of machine learning, and it has distinct advantages over traditional machine learning methods, such as being capable of working on huge volumes of data in complex networks. Furthermore, reinforcement learning also is a type of machine learning, and the main aim of reinforcement learning is to train an agent to carry out a certain task within an uncertain environment. Deep learning and reinforcement learning approaches can be investigated and inspected to be one of the candidate’s algorithms for resource allocation and channel estimation in NOMA system. In this thesis, simulation results clearly indicates that deep learning and reinforcement learning algorithms can provide a superior improvement in terms of diverse performance metrics when Rayleigh and Rician fading channels are considered. Also, in this thesis, a benchmark schemes are also simulated to highlight how much enhancement has been achieved in the system performance when our proposed machine learning models are applied compared to the results obtained by the benchmark schemes.