Adaptive vehicular networking with Deep Learning

Vehicular networks have been identified as a key enabler for future smart traffic applications aiming to improve on-road safety, increase road traffic efficiency, or provide advanced infotainment services to improve on-board comfort. However, the requirements of smart traffic applications also place demands on vehicular networks' quality in terms of high data rates, low latency, and reliability, while simultaneously meeting the challenges of sustainability, green network development goals and energy efficiency. The advances in vehicular communication technologies combined with the peculiar characteristics of vehicular networks have brought challenges to traditional networking solutions designed around fixed parameters using complex mathematical optimisation. These challenges necessitate greater intelligence to be embedded in vehicular networks to realise adaptive network optimisation. As such, one promising solution is the use of Machine Learning (ML) algorithms to extract hidden patterns from collected data thus formulating adaptive network optimisation solutions with strong generalisation capabilities. In this thesis, an overview of the underlying technologies, applications, and characteristics of vehicular networks is presented, followed by the motivation of using ML and a general introduction of ML background. Additionally, a literature review of ML applications in vehicular networks is also presented drawing on the state-of-the-art of ML technology adoption. Three key challenging research topics have been identified centred around network optimisation and ML deployment aspects. The first research question and contribution focus on mobile Handover (HO) optimisation as vehicles pass between base stations; a Deep Reinforcement Learning (DRL) handover algorithm is proposed and evaluated against the currently deployed method. Simulation results suggest that the proposed algorithm can guarantee optimal HO decision in a realistic simulation setup. The second contribution explores distributed radio resource management optimisation. Two versions of a Federated Learning (FL) enhanced DRL algorithm are proposed and evaluated against other state-of-the-art ML solutions. Simulation results suggest that the proposed solution outperformed other benchmarks in overall resource utilisation efficiency, especially in generalisation scenarios. The third contribution looks at energy efficiency optimisation on the network side considering a backdrop of sustainability and green networking. A cell switching algorithm was developed based on a Graph Neural Network (GNN) model and the proposed energy efficiency scheme is able to achieve almost 95% of the metric normalised energy efficiency compared against the "ideal" optimal energy efficiency benchmark and is capable of being applied in many more general network configurations compared with the state-of-the-art ML benchmark.