Machine Learning and Optimization for Resource-Constrained Platforms

Abstract

Artificial intelligence (AI) and machine learning (ML) have been growing at an incredible rate in recent years and they show no sign of stopping. Manufacturing, educational systems, transportation architecture, and genetic research are industries where artificial intelligence algorithms have been developed and found practical applications in which they can increase task efficiency and reduce cost through process optimization, pattern recognition, and automation. At NASA, one of the goals of the cognitive communications project has been to find applications for such algorithms to next-generation communication systems. The goal of this effort is to identify areas and approaches to intelligent system design and implementation which could allow NASA to support a larger space-and ground-based network while simultaneously reducing the operational costs involved with maintaining such a system This paper will evaluate the state of various approaches by searching for algorithms which are feasible to deploy directly onto future space systems with improved processing requirements. We begin by describing a set of heuristics through which algorithms may be compared, emphasizing memory and computational requirements, and heuristic bounds. We then evaluate general-purpose processing platforms onto which such algorithms may be deployed. We also evaluate how such systems may be packaged so as to offer a deterministic set of performance and decision metrics, to make the devices easier for system designers to include in present and future systems. We conclude the paper with a discussion of our findings, as well as where and how this study might continue in the future.