Computationally aware decision making using retrospective intent
Abstract
We perceive the need to enable an autonomous system to minimize bandwidth usage when adapting effectively to its dynamic environment. Minimizing bandwidth demands effective prediction of the potential impacts that may result from a given control signal. Such effective prediction requires high fidelity models for the targets’ future motion. We will develop models to reflect the potential intent of the targets and to capture an understanding of their mechanisms for enacting that intent (e.g., route planning across a road network). We will do this by capitalizing on our extensive experience in using and successfully applying particle filters. Based on this experience, we do not expect a naïve application of textbook particle filters to be effective. More specifically, given that we are attempting to estimate intent, we anticipate that current data will be heavily informative about previous intent. Particle filters struggle when using current data to refine uncertainty about historic events. However, Sequential Monte Carlo (SMC) samplers can be configured to operate in a fixed lag context and offer potential improvements in performance relative to alternative algorithms. We therefore propose to develop Fixed Lag (FL) SMC samplers such that they can process data in the context of the aforementioned models of intent.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 14, 2022
- Source ID
- FA95501917047
Entities
People
- Simon Maskell
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Liverpool