Decision Making in Heterogenous Multi-Agent Systems with Misinformation and Sparse Communication
Abstract
Approved for Public Release: The objective of the proposed work is to establish the theoretical and algorithmic foundations for achi eving resilient multi-agent decision making in the face of misinformation. We consider scenarios like coordinated mapping, surveilla nce, consensus, and long-horizon planning problems where agents must exchange information with one another to achieve coordination a nd autonomous decision making. The presence of misinformation in the system can easily compromise the coordination and invalidate pe rformance guarantees for these systems. Misinformation is the presence of corrupted data in the system arising due to a large span o f causes, from erroneous sensor readings, to unreliable information exchange, to intentional malicious action. These threats to the trustworthiness of exchanged data are unfortunately commonplace in real-world scenarios; particularly in environments that are adver sarial, highly uncertain, and communication constrained as is often the case for Naval missions. Moreover due to the diverse set of potential causes of misinformation, the effects of misinformation on decision making are difficult to model, analyze, and bound. Our starting point to approach this problem is to employ an abstract concept of misinformation as the likelihood of a piece of informat ion (or communication link) being compromised. We refer to this likelihood as the trustworthiness of the information. While modeling the cause of a compromised link requires exhaustive enumeration to capture all possibilities, determining the likelihood that a pie ce of information is compromised is often feasible by comparing against other sources of information in the system. Additionally, ne ighboring observations in the network and intermittent communication opportunities can be exploited to further characterize or learn the presence of misinformation in the system. We capitalize upon this intuition to derive a framework for handling misinformation t hat allows for the application of powerful tools of analysis from probability theory, optimization, and network theory. Under this f ormulation, our objective is to 1) quantify, 2) mitigate, and 3) learn the effects of misinformation on decision making. We target t he recovery of performance guarantees such as convergence, bounded deviation, and convergence rate for autonomous decision making in multi-agent systems with misinformation. We propose to derive a framework for probabilistically bounding or eliminating the effect of misinformation for many important decision making tasks including long-horizon decision making, coverage, surveil-lance, and othe rs that would be highly vulnerable otherwise. Success of our objectives would constitute an important step towards establishing the foundations of reliable autonomous decision-making in the face of misinformation and sparse communication. By enabling the trustwort hiness of autonomous operations in environments prone to misinformation namely ad-hoc, potentially adversarial environments with l imited communications this work will address areas that are core to Naval operations and the Science of Autonomy program.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 07, 2021
- Source ID
- N000142112714
Entities
People
- Stephanie Gil
Organizations
- Office of Naval Research
- President and Fellows of Harvard College
- United States Navy