Proactive edge caching has been regarded as an effective approach to satisfy user experience in mobile networks by providing seamless content transmission and reducing network delay. This is particularly useful in rapidly changing vehicular networks. This paper addresses the proactive edge caching (at roadside unit (RSU)) problem in vehicular networks by mobility prediction, i.e., next RSU prediction. Specifically, the paper proposes a \textit{Hybrid cMAB Proactive Caching System} that implements two parallel online reinforcement learning-based mobility prediction algorithms and allows RSUs to adaptively finalize their predictions to identify as proactive caching nodes. The two parallel prediction algorithms are based on Contextual Multi-armed bandit (cMAB) learning, called Dual-context cMAB and Single-context cMAB. The hybrid system is further developed into two variants: Vehicle-Centric and RSU-Centric. In addition, the paper also conducts comprehensive simulation experiments to evaluate the prediction performance of the proposed hybrid system. They include three traffic scenarios: Commuting traffic, Random traffic and Mixed traffic in Las Vegas, USA and Manchester, UK. With the different road layouts in the two urban areas, the paper aims to generalize the application of the system. Simulation results show that the hybrid Vehicle-Centric system can reach nearly 95% cumulative prediction accuracy in the Commuting traffic scenario and outperform the other methods used for comparison by reaching nearly 80% accuracy in Mixed traffic scenario. Even in the completely Random traffic scenario, it also guarantees a minimum accuracy of nearly 60%.

Proactively caching content at the network edge is particularly effective in high-mobility vehicular networks, where intermittent connection is the major challenge for seamless content transmission. The objective of this paper is to achieve proactive caching in vehicular networks by mobility prediction, specifically by predicting the next roadside unit (RSU) for a vehicle with reinforcement learning techniques. The paper proposes two proactive caching algorithms based on multi-armed bandit (MAB) learning, non-contextual MAB and contextual MAB, respectively. This paper fills the void in the literature regarding the application of MAB learning to mobility-prediction based proactive caching. Their feasibility, superiority, and applicability are evaluated with simulation in two modern cities: Las Vegas, USA with a grid road layout, and Manchester, UK with a more historical layout. Additionally, this paper is the first that proposes to investigate the uncertainty associated with proactive caching systems in the form of entropy with a specifically extended Subjective Logic framework, in order to provide an insight into the underlying link between prediction accuracy and uncertainty.