Active, Assured and Private Learning for Autonomous Networks
Abstract
With the significant advances in autonomous systems, artificial intelligence (AI), communications, sensing and computing, the potent,ial for delivering an actuated, intelligent sensing, and communication network is high. However, given recent failures of purely au,tonomous systems employing AI black boxes, there is a need to invest major effort into the incorporation of risk and cost mitigation, mechanisms into the integration of AI and network design. The notion of risk can encompass a variety of functions including safety,, cost and hard constraints. The fact that engineered systems can control how information is acquired (observations sampled) in a da,ta-driven manner will be generously exploited. We believe that the research proposed herein will be instrumental in the realization,of these low-risk and intelligent networks. We seek to find solutions to the constraint-satisfying and high-performance operation of, a variety of naval relevant systems.To achieve autonomy and desired situational awareness -- essential DoD capabilities -- future m,ulti-agent systems must seamlessly incorporate AI and machine learning (ML) techniques in order to be resilient against uncertainty,and unknown dynamics across multiple scales. However, most ML techniques have been developed for applications and scenarios where th,e process of data collection is not in our control. In contrast, most engineering applications of interest come with the ability to,control and optimize the data collection process. Moreover, in complex environments, the costs and rewards used to direct the data c,ollection are often hard to specify they can be unknown or mis-specified. These challenges are further exacerbated when multiple i,nteracting systems are involved in dynamic environments. Thus, there is a significant need to develop new AI/ML techniques that care,fully address the process of data collection by considering 1) active sampling for learning unknown system dynamics and control poli,cies, 2) cost/reward free methods to provide assured learning, and 3) secure and private multi-agent learning. For this, we have ass,embled a team with multi-disciplinary expertise in AI, information theory, sequential methods, active learning and control.The propo,sed research is of direct relevance to a number of the ONR Science of Autonomy Tough Problems. Our proposal addresses the following,issues: agile methods for safe and robust autonomy including safe learning and use of AI in uncertain environments; as well as a pr,incipled investigation of uncertainty and model representation; as well as planning when beliefs of control systems and environment,s are time-varying or unknown. We seek to learn and extend models at the minimum time scale while providing provable performance gu,arantees. In particular, we focus on finite horizon methods (transient) that achieve good asymptotic performance. Several thrusts,attack the issue of the integration of control and novel computational methods offered by AI/ML/NN with a focus on non-convex proble,ms. Finally, scalable and robust distributed collaboration is essential and addressed through the consideration of both cooperative, and strategic multi-agent systems with imperfect system knowledge.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 01, 2022
- Source ID
- N000142212363
Entities
People
- Urbashi Mitra
Organizations
- Office of Naval Research
- United States Navy
- University of Southern California