Computation and Statistics in High Dimensional Problems of Autonomy
Abstract
Emerging autonomous systems are flushed with massive amounts of data and equipped with high-performance supercomputing resources. They will have to handle massive-scale systems involving tens of thousands of individual components. Examples include the fleets of self-driving cars, delivery drones, small satellite constellations, robotic manufacturing systems, each involving thousands of autonomous and semi-autonomous vehicles interacting with humans and human-operated components. Development of algorithms for autonomy, specifically inference, estimation, control, and learning algorithms, that understand, monitor, and coordinate complex systems at this massive scale must overcome technical challenges in high-dimensional computing and analysis. The main research objective of the proposed research effort is to develop the foundations for new algorithmic and analysis methods to address high-dimensional problems of autonomy. While most existing methods for algorithmic autonomy are plagued by the curse of dimensionality, the new approaches will scale to problems with unprecedented dimensionality by carefully utilizing and further developing tools in compressed computing. We develop algorithms for autonomy that use continuous compressed representations. The new representations are continuous, in the sense that they are in stored directly in a parametric functional form, for instance, instead of point-based grid-like structures that are reminiscent of discretization for ease of computation and add unnecessary complexity to the representations. The new representations are also compressed. The compression is achieved by utilizing novel continuous versions of the tensor decompositions, and they can fit in storage space that is orders of magnitude smaller when compared to naive representations, for instance those based on the regular grid. Yet, at the same time, the new compressed representations still provably comprehend almost the same amount of information. In addition, they can be computationally manipulated efficiently, just like the naive representations such as the regular grid. In this manner, the new representations pack more substantially more information into unit storage space, while preserving the efficiency of computability that existing methods provide. Building autonomy algorithms with these representations leads to orders of magnitude savings in computational resources, especially for extremely-high-dimensional problems of autonomy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 24, 2019
- Source ID
- W911NF1910322
Entities
People
- Sertac Karaman
Organizations
- Army Contracting Command
- Massachusetts Institute of Technology
- United States Army