Data-DRiven Online Scalable SItuation Learning (DDROSSIL)
Abstract
Objective: Modern soldiers are equipped with multimodal sensing units that provide in real-time signal streams of brain waves (EEG), heart activity (EKG), movement (via accelerometers and gyroscopes) and a variety of other physiological signals. These signals contain essential information about what different units experience in the battlefield, which if extracted correctly, can lead to accurate situational awareness and effective decision making. Thus, it is essential to design and analyze effective fusion and learning algorithms that can extract such information. The main objectives of this project are to: i) fuse effectively different types of heterogeneous and nonlinear data streams, while learning their information content about different events in a data-driven fashion; ii) extract pertinent features for accurate clustering and classification of different signal patterns while reducing data dimension; and iii) facilitate real-time and scalable processing. Methods: Our novel data-driven online scalable situation learning framework (DDROSSIL) will entail data-driven mechanisms that perform in real-time scalable signal pattern clustering that can further facilitate reliable tactical decisions. The utilization of proper functions measuring distance from a block diagonal structure and strength of eigenvalues will pave the way to exploit fundamental structural properties that graph similarity matrices need to satisfy for efficient data clustering. Via the utilization of reinforcement learning we devise a framework to determine an optimal data similarity metric construction policy that will boost the performance of clustering in online data acquisition settings and time-varying statistics. The utilization of cross-similarities will enable the recursive processing of small batches of data leading to scalable solutions, while exponential weighing formulations will be used to achieve dynamic graph similarity learning in nonstationary environments. We will design optimal graph filters, which will enable the joint fusion of similar signal streams from different ground units and they will exploit the learnt graph topology information to extract more accurate features and achieve more effective data compression, while respecting power and bandwidth constraints. Significance: DDROSSIL will provide a strong foundation towards an artificial intelligence framework that will help future warfighters to make real-time tactical decisions in harsh environments where statistical models are not available. Relying on heterogenous physiological signals DDROSSIL will perform data-driven clustering that will facilitate real-time situational awareness and accurate decision making under stringent power and bandwidth constraints. The proposed research will introduce benefits in a wide span of fields including data mining, machine learning, as well as remote sensing and medical sensing. The PI will develop a new, research oriented graduate-level course on machine learning and data mining tools, and the research findings of this project will be integrated into the syllabus of new course.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110231
Entities
People
- Ioannis D. Schizas
Organizations
- Army Contracting Command
- United States Army
- University of Texas at Arlington