Strategic Deployment of Artificial Intelligence-Enhanced Cloudlets for Low-latency Human-to-Machine Applications

Abstract

The genesis of mobile cloud computing technology is one of the most significant technical advents of the last decade which can be seen as a marriage between cloud computing and mobile computing technologies. This technical paradigm brings mobile users, telecommunication network operators, and cloud service providers to a common playground, thus providing business opportunities for network operators and cloud service providers. The extension of this facility towards access networks by aggregation of edge-intelligence nodes like cloudlets is one more step forward. A cloudlet is a ``data centre in a box" with enhanced mobility support to bring the cloud closer to mobile users and uses virtual machine abstraction for dynamic resource allocation to trusted mobile users, isolate untrusted mobile users, and support a wide variety of applications without being limited by their process structures, programming languages, or operating systems. To fulfil the ravenous demand for computational resources entangled with the crisp latency requirements of various computationally intensive and mission-critical applications related to augmented reality, autonomous transport, cognitive assistance, and Tactile Internet, installation of cloudlets near access seems to be a very promising solution because of its support for wide geographical network distribution, low latency, mobility and heterogeneity. Finding the optimal cost of cloudlet deployment over urban, suburban, and rural deployment areas with an existing access network, essentially implies finding the optimal placement locations of the cloudlets over the entire deployment area and the optimal amount of computational and storage resources per cloudlet. Technically, this research question leads to an assignment problem, where we need to find the optimal interconnections between mobile devices and cloudlets. In this research, we propose a hybrid cost-optimal cloudlet placement framework over existing fibre-wireless access networks based on mixed-integer non-linear programming. We primarily focus on static cloudlet network planning and placement, i.e., identification of exact optimal cloudlet placement locations over urban, suburban and rural deployment scenarios to provide guidance on the installation cost and assess the workload distribution among different cloudlets and the percentage of incremental energy arising from the presence of cloudlets in the fibre-wireless access networks. Howbeit, we observed that mixed-integer programming based frameworks suffer from scalability issues with large networks and become completely useless when the network data is unavailable. Thus, to overcome this issue, we design analytical frameworks that can provide a quick first-hand estimation of cloudlet deployment cost depending on mobile user density, network architecture, and QoS requirements. We verify that the results produced by this method can be considered as tight lower bounds to that produced by integer programming based frameworks for most practical scenarios. We further perform a parametric analysis to understand the dependence of cloudlet deployment cost on various network parameters. However, depending on the mobility pattern and dynamically varying computational requirements of associated mobile devices, cloudlets at different parts of the network become either overloaded or under-loaded. Thus, we propose an economic and non-cooperative load balancing game for low-latency applications among neighbouring cloudlets, from same as well as different service providers. While addressing load balancing problems, most authors usually stress on minimising the end-to-end latency and do not consider the heterogeneity of neighbouring cloudlets. Nonetheless, in practice, if the job requests are processed within their requested QoS latency target, mobile users should be satisfied. Therefore, instead of formulating a conventional latency minimisation game, we propose a novel utility maximisation game to capture the multi-party economic interaction among heterogeneous neighbouring cloudlets. In this load balancing game, the participating cloudlets achieve their maximum utility when the end-to-end latency is equal to the QoS latency target. With this game formulation, each cloudlet is always interested in receiving some extra job requests and the associated incentives from their neighbouring cloudlets to push their utility towards the maximum point. To implement this game-theoretic load balancing framework, firstly, we propose a centralised mechanism where all the competing cloudlets send their predicted job request arrival rates to a neutral mediator. The mediator computes the Nash equilibrium load balancing strategies for the cloudlets and broadcasts to them before the actual job request arrival. This centralised mechanism also ensures that competing cloudlets are truthful while revealing private information e.g., total incoming job requests. Secondly, we propose a continuous-action reinforcement learning automata-based scheme, which allows each cloudlet to independently compute the Nash equilibrium in a completely distributed network setting. We critically study the convergence properties of the designed learning algorithm, scaffolding our understanding of the underlying load balancing game for faster convergence, and study the impacts of exploration and exploitation on learning accuracy. After investigating the cloudlet placement and load balancing problems, we investigate the role of edge-intelligence servers like cloudlets in deploying low-latency human-to-machine applications like teleoperation, immersive virtual/augmented reality, and industrial automotive control over long-distance access networks. Such applications are being realised through Tactile Internet that allows users to control remote things and involve the bi-directional transmission of video, audio, and haptic data. However, the end-to-end propagation latency presents a stubborn bottleneck, which can be alleviated by using various artificial intelligence-based application layer and network layer prediction algorithms, e.g., forecasting and preempting haptic feedback transmission. To gain proper insights, we study the experimental data on traffic characteristics of control signals and haptic feedback samples obtained through virtual reality-based human-to-machine teleoperation. Moreover, we propose the installation of edge-intelligence servers between master and slave devices to implement the preemption of haptic feedback from control signals. Harnessing virtual reality-based teleoperation experiments, we further propose a two-stage artificial intelligence-based module for forecasting haptic feedback samples. The first-stage unit is a supervised binary classifier that detects if haptic sample forecasting is necessary and the second-stage unit is a guided reinforcement learning unit that ensures haptic feedback samples are forecasted accurately when different types of material are present. Furthermore, by evaluating analytical expressions, we show the feasibility of deploying remote human-to-machine teleoperation over fibre backhaul by using our proposed artificial intelligence-based module, even under heavy traffic intensity.