Distributed Communication Optimization for Mission-Driven Wireless Networks
Abstract
Approved for public release.Today#s wireless networks are increasingly designed to support applications with strict performance requirements, e.g., in order to maintain awareness of the surrounding environment for applications like autonomous vehicles or augmented reality. However, despite recent advances in wireless technology, significant network infrastructure is often required to meet thenominal performance requirements of these applications. In remote or resource-challenged environments, such infrastructure generally cannot be assumed. Thus, supporting communication intensive applications in such environments raises significant research challenges.In this project, we aim to enable mission-driven wireless networks that relax rigid application performance specifications and instead allow applications to adapt to mission needs and current network availability. This adaptation comes from the insight that while strict network quality-of-service requirements are easy to specify, they are not always necessary to ensure satisfactory application performance. Indeed, in naval missions, a commander might care much more about whether there is #enough# information to make decisions, not whether the maximum amount of information has been sent over a network. Defining #enough,# however, is difficult: while some applications like distributed machine learning or video streaming might have relatively well-defined performance standards, even these may be subjective, with different users having different expectations depending on the specific context in which the application is used. Some applications may not even have well-defined models of the quality of their users# experience. Without such models, it is difficult to adapt applications# network usage to current network conditions.We address the above challenges by designing online learning mechanisms that allow the network and wireless users to learn how to best adapt applications# network demands to current network conditions. We first adapt the traditional idea of using virtual #prices# to signal network congestion levels are at any given time. By carefully calibrating the prices to users# reactions, we can then ensure that users moderate their demands according to the available network resources. Moreover, to limit the feedback required from users, we leverage #side information# collected from shorter-timescale network demands and integrate this information into our learning algorithms. We then consider distributed applications in which different users# application utilities are coupled together, e.g., if multiple sensors are collectively sensing an environment. Optimally trading off between application utility and data transmissions then requires coordination between these users, which is particularly difficult when communication is limited. Thus, we formulate multi-agent learning algorithms that allow usersto learn how to coordinate under limited communication. We finally integrate our virtual pricing and user coordination algorithms with algorithms that learn neural network approximators for application utility functions, allowing for more expressive and general types of application utilities. We finally plan to validate our algorithms in terms of theoretical guarantees on their convergence and optimality, as well as simulation-based validation with network traces and simulators. We will use testbeds with Raspberry Pi devices communicating over Wi-Fi and LoRA Feather Radio devices to experimentally validate our algorithm performance for two example applications: distributed model training and environmental sensing.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 24, 2024
- Source ID
- N000142412073
Entities
People
- Carlee Joe-Wong
Organizations
- Carnegie Mellon University
- Office of Naval Research
- United States Navy