Thrust 1

Reconfigurable Framework for Integrated Command and Control

Lead PI: Manu Sridharan

This research thrust aims to develop a new software plane and functional architecture, guided by robust machine learning models, to assess dynamic mission requirements, available resources, and environmental factors.  We are working on the adaptation of machine learning (ML) models to a distributed environment. Consider that we have a set of local nodes communicating with a central station. The central station has a global ML model for a certain task, which is replicated at the local nodes. Each node can update these models based on their own observations. The global model needs to be updated based on these local updates. However, the communication links between the local nodes and the central station are dynamic and of varying quality. Transmission of model parameters may be affected by these uncertain communication links, and may not be received synchronously at the central station. This problem falls within the domain of Federated Learning (FL); however, there is little work on how FL works with unreliable communication settings. Going forward, we will consider a more general setting which is aligned with FNC3, wherein one can imagine peer-to-peer communications for enabling local model reconciliations, and using such reconciled models to represent the local model (encompassing a plurality of nodes in proximity).

Leveraging the ML models as inputs, the second component of this thrust will develop foundational techniques to automatically verify and synthesize a module configuration for a mission, given a specification of higher-level mission goals.  The PIs will develop synthesis techniques that both guarantee correctness and minimize usage of network resources, ensuring maximum flexibility in deployment.  The work will initially focus on bandwidth usage verification, as excessive bandwidth usage can compromise network performance and overwhelm servers.  The PIs aim to discover bandwidth usage bugs in existing applications, compute network-wide bandwidth usage bounds, and eventually synthesize configurations guaranteed to adhere to bounds on bandwidth and other resources.