Translating Commander Intent into Network Policies: Towards Autonomous Wireless Networks
Abstract
Wireless networks increasing capacity and flexibility have fueled a range of mobile applicationsthat promise increasingly comprehensive, real-time interactions between users and theirenvironments, ranging from augmented reality (AR) systems that overlay digital content onto theuser s field of view to distributed sensors that collect and analyze data in real time. While theseapplications are useful in naval scenarios, e.g., to identify and label threats, they can requireintensive, high-throughput communication. To meet these needs, network policies should carefullymanage which applications receive access to network resources, which are often limited in navalenvironments. For instance, analyzing some sensor data streams may be more urgent than others.Translating the needs of an ongoing naval mission into network policies is difficult: navalcommanders will likely not have the time to produce detailed specifications of how to prioritizedifferent applications, particularly if these priorities change with on-the-ground conditions. Evenif such specifications exist, the dynamic nature of wireless networks can make it difficult todetermine which network policies best satisfy them. The proposed work addresses these technicalchallenges by designing online learning algorithms that can (i) infer commander intent fromlimited feedback and (ii) intelligently sense the network state to find network policies satisfyingthe inferred intent. The main anticipated outcomes of this project are the design; softwareimplementations; and theoretical, simulation-based, and experimental evaluations of our learningalgorithms on both synthetic and realistic network performance data.We design these online learning algorithms following three related research directions:(1) First, we will design algorithms that learn when to sense the network state given commanderintent, accounting for the cost of this sensing and potential inaccuracies in the sensing results. Thetwo main research tasks are to account for (i) potential inaccuracies in the sensed observations aswell as (ii) temporal correlations between the network state.(2) We will then design algorithms to infer the network policies that best satisfy unspecifiedcommander intent. The two main research tasks are (i) learning application priorities from binaryfeedback on whether the incurred performance was satisfactory and (ii) analyzing the rate ofconvergence when this feedback may be delayed or missing.(3) Finally, we will (i) integrate our network sensing and intent inference algorithms and (ii)develop extensions that scale to larger networks with general topologies.We will evaluate the algorithms developed by theoretically analyzing their convergence rates andoptimality relative to heuristic baselines. We will also evaluate them numerically on data fromWiFi and LoRA network simulators and implement them on a small wireless testbed.By creating a framework for autonomous wireless network management, the proposed work willsignificantly increase the ability of Department of Defense communication networks to respond toreal-time mission needs by intelligently prioritizing applications that compete for the samenetwork resources. Such autonomous network management may enable soldiers to useincreasingly demanding applications such as AR, increasing their effectiveness on the battlefield.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 02, 2021
- Source ID
- N000142112128
Entities
People
- Carlee Joe-Wong
Organizations
- Carnegie Mellon University
- Office of Naval Research
- United States Navy