THIS IS A CONTINUATION OF N00014-12-1-1000 persistent Decentralized Online Tasks (pDOT): An Online Optimization Approach to Multi-Agent Persistent Monitoring in Uncertain Environments
Abstract
Statement of Work:The research effort will bridge the currently disparate fields of machine learning, control theory, information theory, and online discrete algorithms to develop new algorithms with provable guarantees for decentralized online optimization that can deliver real-time performance for autonomous teams.Objective:The persistent Decentralized Online Tasks (pDOT) project is a basic research program that will achieve thebreakthroughs in decentralized online optimization necessary to effectively utilize the Navy s arsenal of autonomous vehicles with minimal human intervention. The research effort will bridge the currently disparate fields of machine learning, control theory, information theory, and online discrete algorithms to develop new algorithms with provable guarantees for decentralized online optimization that can deliver real-time performance for autonomous teams. The PI will consider a unifying persistent surveillance scenario in which the pDOT agents will be a group of naval vehicles (unmanned aerial, ground, and surface vehicles), tasked with monitoring a maritime region and targets within the region for on-line detection of normal and abnormal activity patterns over long periods of time.Approach:The research will seek to endow a group of pDOT agents with the ability to learn, estimate, detect events, and act in a timely fashion in response to online changes in a data rich environment. The PI proposes to develop decentralized online optimization algorithms with performance guarantees to enable reliable learning, estimation, and control tasks in data-rich environments. An important objective is to characterize the solution quality and the timeliness of computation, so that the agents can be trusted to perform their tasks while ensuring the safety of people and goods in the port. In order to develop algorithmic and analysis tools suitable for decentralized online optimization, the PI will focus their efforts in this project in three thrusts. Thrust P will develop decentralized algorithms for Persistence in Dynamic Environments with Online Data. Thrust M will develop algorithms for acquiring decentralized online Probabilistic Models under Perceptual Uncertainty. Thrust I will address the Integration of Persistent Monitoring with Probabilistic Models and consider the technical and validations issues when moving from theory to experiment.Overall Merit and ONR Mission/Relevance:The research is innovatrive in that it includes the development of rigorous algorithmic-analysis techniques for decentralized online optimization that provides performance predictions through sound estimates of the closeness a decentralized online solution to a centralized, global offline optimal solution; such performance guarantees contribute tremendously to trust in these systems, which is critical for their deployment. The Navy is moving towards deploying large, complex systems that are beyond centralized control. A canonicalexample of such a system is a fleet of unmanned vehicles with limited communications operating in a dynamic environment. Important characteristics of these systems are that 1) they are decentralized (i.e., system components can take independent actions), and 2) the environment in which the system operates is not necessarily known a priori, and is revealed over time. The objective of this topic is to develop scientific principles and algorithms for solving decentralized, online optimization problemsProgress:We last reported on our algorithm that takes advantage of spatial correlation in attacking the multirobot persistentmonitoring problem. Intuitively, two points in space that are geometrically close should have similar "properties" of interest, due to spatial continuity. Hence, a pDOT agent may only need to visit one of these two sites to get sufficient information about both. We proposed a novel non-linear extension to the classical orienteering problem (OP), called the Correlated Orienteering Problem (COP). We u
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 23, 2016
- Source ID
- N000141612787
Entities
People
- Daniela L. Rus
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy