With the proliferation of mobile computing and Internet-of-Things (IoT), billions of mobile and IoT devices are connected to the Internet, generating zillions of Bytes of data at the network edge. Driving by this trend, there is an urgent need to push the artificial intelligence (AI) frontiers to the network edge to unleash the potential of the edge big data fully. This dissertation aims to comprehensively study collaborative learning and optimization algorithms to build a foundation of edge intelligence. Under this common theme, this dissertation is broadly organized into three parts. The first part of this study focuses on model learning with limited data and limited computing capability at the network edge. A global model initialization is first obtained by running federated learning (FL) across many edge devices, based on which a semi-supervised algorithm is devised for an edge device to carry out quick adaptation, aiming to address the insufficiency of labeled data and to learn a personalized model efficiently. In the second part of this study, collaborative learning between the edge and the cloud is studied to achieve real-time edge intelligence. More specifically, a distributionally robust optimization (DRO) approach is proposed to enable the synergy between local data processing and cloud knowledge transfer. Two attractive uncertainty models are investigated corresponding to the cloud knowledge transfer: the distribution uncertainty set based on the cloud data distribution and the prior distribution of the edge model conditioned on the cloud model. Collaborative learning algorithms are developed along this line. The final part focuses on developing an offline model-based safe Inverse Reinforcement Learning (IRL) algorithm for connected Autonomous Vehicles (AVs). A reward penalty is introduced to penalize unsafe states, and a risk-measure-based approach is proposed to mitigate the model uncertainty introduced by offline training. The experimental results demonstrate the improvement of the proposed algorithm over the existing baselines in terms of cumulative rewards.

In recent years, Artificial Intelligence (AI) (e.g., Deep Neural Networks (DNNs), Transformer) has shown great success in real-world applications due to its superior performance in various cognitive tasks. The impressive performance achieved by AI models normally accompanies the cost of enormous model size and high computational complexity, which significantly hampers their implementation on resource-limited Cyber-Physical Systems (CPS), Internet-of-Things (IoT), or Edge systems due to their tightly constrained energy, computing, size, and memory budget. Thus, the urgent demand for enhancing the \textbf{Efficiency} of DNN has drawn significant research interests across various communities. Motivated by the aforementioned concerns, this doctoral research has been mainly focusing on Enabling Deep Learning at Edge: From Efficient and Dynamic Inference to On-Device Learning. Specifically, from the inference perspective, this dissertation begins by investigating a hardware-friendly model compression method that effectively reduces the size of AI model while simultaneously achieving improved speed on edge devices. Additionally, due to the fact that diverse resource constraints of different edge devices, this dissertation further explores dynamic inference, which allows for real-time tuning of inference model size, computation, and latency to accommodate the limitations of each edge device. Regarding efficient on-device learning, this dissertation starts by analyzing memory usage during transfer learning training. Based on this analysis, a novel framework called "Reprogramming Network'' (Rep-Net) is introduced that offers a fresh perspective on the on-device transfer learning problem. The Rep-Net enables on-device transferlearning by directly learning to reprogram the intermediate features of a pre-trained model. Lastly, this dissertation studies an efficient continual learning algorithm that facilitates learning multiple tasks without the risk of forgetting previously acquired knowledge. In practice, through the exploration of task correlation, an interesting phenomenon is observed that the intermediate features are highly correlated between tasks with the self-supervised pre-trained model. Building upon this observation, a novel approach called progressive task-correlated layer freezing is proposed to gradually freeze a subset of layers with the highest correlation ratios for each task leading to training efficiency.

Voltage Source Converters (VSCs) have been widely used in grid-connected applications with Distributed Energy Resource (DER) and Electric Vehicle (EV) applications. Replacement of traditional thyristors with Silicon/Silicon-Carbide based active switches provides full control capability to the converters and allows bidirectional power flow between the source and active loads. In this study, advanced control strategies for DER inverters and EV traction inverters will be explored.Chapter 1 gives a brief introduction to State-of-the-Art of VSC control strategies and summarizes the existing challenges in different applications. Chapter 2 presents multiple advanced control strategies of grid-connected DER inverters. Various grid support functions have been implemented in simulations and hardware experiments under both normal and abnormal operating conditions. Chapter 3 proposes an automated design and optimization process of a robust H-infinity controller to address the stability issue of grid-connected inverters caused by grid impedance variation. The principle of the controller synthesis is to select appropriate weighting functions to shape the systems closed-loop transfer function and to achieve robust stability and robust performance. An optimal controller will be selected by using a 2-Dimensional Pareto Front. Chapter 4 proposes a high-performance 4-layer communication architecture to facilitate the control of a large distribution network with high Photovoltaic (PV) penetration. Multiple strategies have been implemented to address the challenges of coordination between communication and system control and between different communication protocols, which leads to a boost in the communication efficiency and makes the architecture highly scalable, adaptive, and robust. Chapter 5 presents the control strategies of a traditional Modular Multilevel Converter (MMC) and a novel Modular Isolated Multilevel Converter (MIMC) in grid-connected and variable speed drive applications. The proposed MIMC is able to achieve great size reduction for the submodule capacitors since the fundamental and double-line frequency voltage ripple has been cancelled. Chapter 6 shows a detailed hardware and controller design for a 48 V Belt-driven Starter Generator (BSG) inverter using automotive gate driver ICs and microcontroller. The inverter prototype has reached a power density of 333 W/inch3, up to 200 A phase current and 600 Hz output frequency.

Edge networks pose unique challenges for machine learning and network management. The primary objective of this dissertation is to study deep learning and adaptive control aspects of edge networks and to address some of the unique challenges therein. This dissertation explores four particular problems of interest at the intersection of edge intelligence, deep learning and network management. The first problem explores the learning of generative models in edge learning setting. Since the learning tasks in similar environments share model similarity, it is plausible to leverage pre-trained generative models from other edge nodes. Appealing to optimal transport theory tailored towards Wasserstein-1 generative adversarial networks, this part aims to develop a framework which systematically optimizes the generative model learning performance using local data at the edge node while exploiting the adaptive coalescence of pre-trained generative models from other nodes. In the second part, a many-to-one wireless architecture for federated learning at the network edge, where multiple edge devices collaboratively train a model using local data, is considered. The unreliable nature of wireless connectivity, togetherwith the constraints in computing resources at edge devices, dictates that the local updates at edge devices should be carefully crafted and compressed to match the wireless communication resources available and should work in concert with the receiver. Therefore, a Stochastic Gradient Descent based bandlimited coordinate descent algorithm is designed for such settings. The third part explores the adaptive traffic engineering algorithms in a dynamic network environment. The ages of traffic measurements exhibit significant variation due to asynchronization and random communication delays between routers and controllers. Inspired by the software defined networking architecture, a controller-assisted distributed routing scheme with recursive link weight reconfigurations, accounting for the impact of measurement ages and routing instability, is devised. The final part focuses on developing a federated learning based framework for traffic reshaping of electric vehicle (EV) charging. The absence of private EV owner information and scattered EV charging data among charging stations motivates the utilization of a federated learning approach. Federated learning algorithms are devised to minimize peak EV charging demand both spatially and temporarily, while maximizing the charging station profit.

Applications over a gesture-based human-computer interface (HCI) require a new user login method with gestures because it does not have traditional input devices. For example, a user may be asked to verify the identity to unlock a device in a mobile or wearable platform, or sign in to a virtual site over a Virtual Reality (VR) or Augmented Reality (AR) headset, where no physical keyboard or touchscreen is available. This dissertation presents a unified user login framework and an identity input method using 3D In-Air-Handwriting (IAHW), where a user can log in to a virtual site by writing a passcode in the air very fast like a signature. The presented research contains multiple tasks that span motion signal modeling, user authentication, user identification, template protection, and a thorough evaluation in both security and usability. The results of this research show around 0.1% to 3% Equal Error Rate (EER) in user authentication in different conditions as well as 93% accuracy in user identification, on a dataset with over 100 users and two types of gesture input devices. Besides, current research in this area is severely limited by the availability of the gesture input device, datasets, and software tools. This study provides an infrastructure for IAHW research with an open-source library and open datasets of more than 100K IAHW hand movement signals. Additionally, the proposed user identity input method can be extended to a general word input method for both English and Chinese using limited training data. Hence, this dissertation can help the research community in both cybersecurity and HCI to explore IAHW as a new direction, and potentially pave the way to practical adoption of such technologies in the future.

Collision-free path planning is also a major challenge in managing unmanned aerial vehicles (UAVs) fleets, especially in uncertain environments. The design of UAV routing policies using multi-agent reinforcement learning has been considered, and propose a Multi-resolution, Multi-agent, Mean-field reinforcement learning algorithm, named 3M-RL, for flight planning, where multiple vehicles need to avoid collisions with each other while moving towards their destinations. In this system, each UAV makes decisions based on local observations, and does not communicate with other UAVs. The algorithm trains a routing policy using an Actor-Critic neural network with multi-resolution observations, including detailed local information and aggregated global information based on mean-field. The algorithm tackles the curse-of-dimensionality problem in multi-agent reinforcement learning and provides a scalable solution. The proposed algorithm is tested in different complex scenarios in both 2D and 3D space and the simulation results show that 3M-RL result in good routing policies. Also as a compliment, dynamic data communications between UAVs and a control center has also been studied, where the control center needs to monitor the safety state of each UAV in the system in real time, where the transition of risk level is simply considered as a Markov process. Given limited communication bandwidth, it is impossible for the control center to communicate with all UAVs at the same time. A dynamic learning problem with limited communication bandwidth is also discussed in this paper where the objective is to minimize the total information entropy in real-time risk level tracking. The simulations also demonstrate that the algorithm outperforms policies such as a Round & Robin policy.

Low frequency oscillations (LFOs) are recognized as one of the most challenging problems in electric grids as they limit power transfer capability and can result in system instability. In recent years, the deployment of phasor measurement units (PMUs) has increased the accessibility to time-synchronized wide-area measurements, which has, in turn, enabledthe effective detection and control of the oscillatory modes of the power system. This work assesses the stability improvements that can be achieved through the coordinated wide-area control of power system stabilizers (PSSs), static VAr compensators (SVCs), and supplementary damping controllers (SDCs) of high voltage DC (HVDC) lines, for damping electromechanical oscillations in a modern power system. The improved damping is achieved by designing different types of coordinated wide-area damping controllers (CWADC) that employ partial state-feedback. The first design methodology uses a linear matrix inequality (LMI)-based mixed H2/Hinfty control that is robust for multiple operating scenarios. To counteract the negative impact of communication failure or missing PMU measurements on the designed control, a scheme to identify the alternate set of feedback signals is proposed. Additionally, the impact of delays on the performance of the control design is investigated. The second approach is motivated by the increasing popularity of artificial intelligence (AI) in enhancing the performance of interconnected power systems. Two different wide-area coordinated control schemes are developed using deep neural networks (DNNs) and deep reinforcement learning (DRL), while accounting for the uncertainties present in the power system. The DNN-CWADC learns to make control decisions using supervised learning; the training dataset consisting of polytopic controllers designed with the help of LMI-based mixed H2/Hinfty optimization. The DRL-CWADC learns to adapt to the system uncertainties based on its continuous interaction with the power system environment by employing an advanced version of the state-of-the-art deep deterministic policy gradient (DDPG) algorithm referred to as bounded exploratory control-based DDPG (BEC-DDPG). The studies performed on a 29 machine, 127 bus equivalent model of theWestern Electricity Coordinating Council (WECC) system-embedded with different types of damping controls have demonstrated the effectiveness and robustness of the proposed CWADCs.

With demand for increased efficiency and smaller carbon footprint, power system operators are striving to improve their modeling, down to the individual consumer device, paving the way for higher production and consumption efficiencies and increased renewable generation without sacrificing system reliability. This dissertation explores two lines of research. The first part looks at stochastic continuous-time power system scheduling, where the goal is to better capture system ramping characteristics to address increased variability and uncertainty. The second part of the dissertation starts by developing aggregate population models for residential Demand Response (DR), focusing on storage devices, Electric Vehicles (EVs), Deferrable Appliances (DAs) and Thermostatically Controlled Loads (TCLs). Further, the characteristics of such a population aggregate are explored, such as the resemblance to energy storage devices, and particular attentions is given to how such aggregate models can be considered approximately convex even if the individual resource model is not. Armed with an approximately convex aggregate model for DR, how to interface it with present day energy markets is explored, looking at directions the market could go towards to better accommodate such devices for the benefit of not only the prosumer itself but the system as a whole.

The first half of this dissertation introduces a minimum cost incentive mechanism for collecting discrete distributed private data for big-data analysis. The goal of an incentive mechanism is to incentivize informative reports and make sure randomization in the reported data does not exceed a target level. It answers two fundamental questions: what is the minimum payment required to incentivize an individual to submit data with quality level $\epsilon$? and what incentive mechanisms can achieve the minimum payment? A lower bound on the minimum amount of payment required for guaranteeing quality level $\epsilon$ is derived. Inspired by the lower bound, our incentive mechanism (WINTALL) first decides a winning answer based on reported data, then pays to individuals whose reported data match the winning answer. The expected payment of WINTALL matches lower bound asymptotically. Real-world experiments on Amazon Mechanical Turk are presented to further illustrate novelty of the principle behind WINTALL.

The second half studies problem of iterative training in Federated Learning. A system with a single parameter server and $M$ client devices is considered for training a predictive learning model with distributed data. The clients communicate with the parameter server using a common wireless channel so each time, only one device can transmit. The training is an iterative process consisting of multiple rounds. Adaptive training is considered where the parameter server decides when to stop/restart a new round, so the problem is formulated as an optimal stopping problem. While this optimal stopping problem is difficult to solve, a modified optimal stopping problem is proposed. Then a low complexity algorithm is introduced to solve the modified problem, which also works for the original problem. Experiments on a real data set shows significant improvements compared with policies collecting a fixed number of updates in each iteration.

The availability of data for monitoring and controlling the electrical grid has increased exponentially over the years in both resolution and quantity leaving a large data footprint. This dissertation is motivated by the need for equivalent representations of grid data in lower-dimensional feature spaces so that machine learning algorithms can be employed for a variety of purposes. To achieve that, without sacrificing the interpretation of the results, the dissertation leverages the physics behind power systems, well-known laws that underlie this man-made infrastructure, and the nature of the underlying stochastic phenomena that define the system operating conditions as the backbone for modeling data from the grid.

The first part of the dissertation introduces a new framework of graph signal processing (GSP) for the power grid, Grid-GSP, and applies it to voltage phasor measurements that characterize the overall system state of the power grid. Concepts from GSP are used in conjunction with known power system models in order to highlight the low-dimensional structure in data and present generative models for voltage phasors measurements. Applications such as identification of graphical communities, network inference, interpolation of missing data, detection of false data injection attacks and data compression are explored wherein Grid-GSP based generative models are used.

The second part of the dissertation develops a model for a joint statistical description of solar photo-voltaic (PV) power and the outdoor temperature which can lead to better management of power generation resources so that electricity demand such as air conditioning and supply from solar power are always matched in the face of stochasticity. The low-rank structure inherent in solar PV power data is used for forecasting and to detect partial-shading type of faults in solar panels.