As mobile devices have become the preferred tool for communication, work, and entertainment, traffic at the edge of the network is growing more rapidly than ever. To improve user experience, commodity servers are deployed in the edge to form a decentralized mobile edge computing (MEC) network of mini datacenters each serving a localized region. This project is aimed to (1) design a scalable and adaptive connectivity architecture for the MEC servers, (2)  deploy this architecture geographically and assign workloads to the servers for maximal offloading benefit, and (3) devise stochastic-modeling based algorithms to cope with time-varying workloads. 

Network latency and packet loss rate in a computer network, social distance in a social network, or in general, any quantity indicating a node's remoteness from one another in a network, examples of network dissimilarity. The Network Dissimilarity Estimation (NDE) problem is: given sample measurements, how to compute the dissimilarity for any pair of nodes without more measurements. The geometric embedding approach to NDE assumes a linear relationship between dissimilarity and its intrinsic metric. We conjecture that this relationship follows a power law. The project is to validate and apply the conjecture to several practical applications. 

Location fingerprinting, a range-free approach to GPS-free localization, requires a prior training set of fingerprints, each sampled at a known location, so that future fingerprints can be located by comparing to this set. Conventionally, the fingerprint space is defined as a feature vector space. In practice, it is hard to find a quality feature space robust to device heterogeneity and environment and infrastructure dynamics. We advocate a fundamentally different model where a fingerprint is defined as a dissimilarity measurement associated with a pair of locations. We propose a localization framework for this model based on Riemannian geometrical embedding.

The mainstream approach to distributed data storage is essentially data-centric, aiming to preserve certain content relationships. We advocate the concept that data are not what their content is, but are what they are queried for. Consequently, we propose an approach that preserves query-centric relationships. In this era of big data, although huge data are generated, only a small subset of them is queried, hence useful. By adopting a query-centric approach, the optimization can focus only on the useful data, for the benefit of the queries.  The project will involve all three crucial components: storage, indexing, and networking. 

We are interested in enabling techniques for target monitoring using low-cost sensor networks. A “sensor” is a small computing device that can read the environment and process information; it can be a mobile phone or an autonomous underwater vehicle with sensing capabilities, or typical sensors as in a conventional sensor network. We want to devise a solution that can tell us whether a special event of interest happens, when and where, quickly, accurately, and efficiently. Our contribution consists of a localization solution based on a novel hierarchical machine learning method using hop-count information only and a publish/subscribe solution allowing for quick and efficient notification when an event occurs.