Toward an intelligent wireless edge : computing, centralized and decentralized learning

Abstract

Recent years have witnessed unprecedented successes of artificial intelligence (AI), which is transforming virtually every branch of science and technology. Meanwhile, with the rapid advancement of wireless communications and the proliferation of Internet of Things (IoT), billions of mobile gadgets and IoT devices are being connected to the Internet, generating huge volume of data at the network edge. Driven by this trend, it is highly desiderated to push the AI frontiers to the network edge so as to fully unleash the potential of IoT big data. To this end, mobile edge computing (MEC) emerges as a promising solution that allows rapid access to the tremendous real-time IoT data for quick AI-model training. The resultant new interdiscipline is called “edge intelligence” or “edge learning”. This dissertation contributes to this emerging area by dealing with several key challenges toward realizing the vision of an intelligent wireless edge, spanning from MEC to centralized and decentralized edge learning. First, we investigate the scheduling of an MEC system with random user arrivals. The long-term energy consumption and latency are minimized by jointly optimizing offloading decision, user selection, and transmit power allocation. This problem is formulated as an infinite-horizon Markov decision process (MDP). To cope with the huge and uncertain dimension of system state space incurred by random user arrivals, a novel low-complexity approximate MDP framework is proposed and validated. Next, building upon MEC, we study efficient wireless communication design for centralized edge learning, and propose a reconfigurable intelligent surface (RIS)-assisted learning-centric wireless communication scheme. In contrast to conventional communication systems where the principal criteria are to maximize throughput, the learning-centric scheme aims at maximizing learning performance. Specifically, we minimize the learning errors of all the participating users by jointly optimizing transmit power of the mobile users, beamforming vectors at the base station, and phase-shift matrix of the RIS. Simulations demonstrate significant gains in terms of learning accuracy benefitting from deploying an RIS and the learning-centric scheme. Last, to ease the severe privacy concerns in centralized edge learning, we further investigate decentralized edge learning, where a novel topology-optimized federated edge learning (TOFEL) scheme is proposed to tackle the heterogeneity issue. A joint topology and computing speed optimization problem is formulated to minimize the latency and energy consumption. Then, a penalty-based method is proposed to solve the mixed integer nonlinear problem. To facilitate real-time decision making, an imitation learning-based method is developed, where deep neural networks imitating the penalty-based method are trained offline, and then deployed for online inference. Simulations show that the proposed TOFEL scheme remarkably accelerates federated learning, and achieves higher energy efficiency. Moreover, to validate the effectiveness of our proposed schemes for close-to-reality learning tasks, we apply the proposed schemes developed in both centralized and decentralized learning systems to 3D object detection of autonomous driving in CARLA simulator. The experimental results demonstrate superior detection accuracies of our proposed schemes over various benchmarks.