Distributed Data Processing and Machine Learning Algorithms for Wireless Battlefield Networks

Edge (fog) computing has recently emerged as a promising paradigm to improve the latency of real-time applications in IoT (Internet of Things) and mobile networks. By using computing resources near IoT devices for local storage and preliminary data processing, fog computing can improve the delay incurred due to the bandwidth bottleneck between devices and the cloud. Adapting such architectures for distributed data processing in a wireless battlefield network poses several challenges that cannot be adequately addressed by todayÕs (cloud or fog) computing and communication models: First, the wireless battlefield network is a harsh environment. Wireless nodes are subject to limited power, limited computation capability, limited channel state information, and intermittent connectivity. Today's distributed computing models are focused on data centers where the communication medium is the data center fabric (a switch network), which is fundamentally different from a wireless battlefield network. Second, military data processing applications require stringent delay guarantees. The resource-constrained nodes in a wireless environment will not be able to rely solely on their own limited resources to perform all their computing needs. Hence nodes need to make decisions about what needs to be computed locally and what needs to be transferred to other nodes in a distributed manner, based on knowledge which is locally available or involves only a limited exchange of messages with their neighbors. The recent advances in fog (edge) computing mainly focus on user-oriented computation assisted by access points, while this project focuses on network-oriented computation where nodes should cooperate to perform a distributed computing job. This project will design efficient communication and resource allocation algorithms to enable fast and distributed data processing in wireless battlefield networks. Specifically, the project will achieve two main objectives: (1) Processing computation graphs in a dispersed wireless computing network: Distributed computing models (e.g., MapReduce, Hadoop, Spark) rely on the representation of jobs as computations graphs, in which each vertex represents a computing task (e.g., a local gradient computation in a machine learning job), and each directed edge represents the data flow from one vertex to a descendant vertex. We first focus on processing computation graphs under an ideal interference graph model of wireless networks. This is considerably harder than distributed processing in data centers, since machines in a data center do not interfere with each other while mobile computing units necessarily do when exchanging information along the edges of computation graphs. The project will create task-to-node assignment and flow scheduling algorithms that can utilize the computation and communication resources efficiently to optimize network-level performance metrics such as job computation latency. Our techniques rely on a combination of linear programming relaxations, greedy scheduling, and distributed convex optimization. (2) Machine learning algorithms with wireless network structure: To apply distributed computing to realistic wireless battlefield networks, we need to design algorithms that can deal with unknown channel statistics and infrequent wireless channel state feedbacks. Machine Learning (ML) algorithms for convex optimization and MAB (Multi Armed Bandit) are natural candidates to study such problems. However, when the structure of wireless networks is imposed on these problems, the ML solutions are either inapplicable (as in the case of primal-dual algorithms which require a known bound on the dual variables) or are computationally infeasible (as in the case of structured MABs which require the solution to a semi-infinite LP at each step). This project will design algorithms that exploit the key ideas from ML, but make the algorithms viable in the context of wireless networks.