Title: TAT-C ML: Machine Learning for Enhanced Trade-space Analysis of Constellations
Presenting Author: Daniel Selva
Organization: Cornell University
Co-Author(s): Nozomi Hitomi1, Cornell University; Philip W. Dabney, NASA GSFC; Veronica Foreman, MIT; Paul Grogan, Stevens Institute of Technology; Matthew P. Holland, NASA GSFC; Steven P. Hughes, NASA GSFC; Sreeja Nag, Bay Area Environmental Research Institute; Michael G. Plante, NASA GSFC; Afreen Siddiqi, MIT; Jacqueline Le Moigne, NASA GSFC

Abstract:
Recent architecture studies on Earth observing systems (ESAS2017, NOAA) have emphasized the importance of distributed satellite missions to meet future observation requirements in a cost-effective way. The tools that we currently use in Pre-Phase and Phase A studies to analyze and explore the space of possible mission architectures are not well-suited to tackle such distributed systems, due to the computational effort required to simulate one such system and the combinatorial explosion of the number of alternatives. As a consequence, the Trade-space Analysis Tool for Constellations (TAT-C) was developed as an AIST14-funded project. The current version of TAT-C couples a basic trade-space exploration tool with fast models to calculate metrics such as coverage, cost and risk. However, TAT-C is still limited in terms of the number and types of trades that it can analyze, as it uses a brute-force design-of-experiments approach to searching the solution space. To address this, one of the goals of the TAT-C ML AIST16 project is to incorporate artificial intelligence and machine learning into TAT-C in order to make the search more efficient. Specifically, a state-of-the-art evolutionary algorithm is used that maintains a pool of traditional domain-independent operators such as crossover and mutation and domain-specific operators that encode expert knowledge (heuristics) to guide the search. Such expert knowledge may be available a priori (e.g., increasing altitude to improve coverage, spreading satellites across more planes to improve tail statistics for revisit times) or it can be learned online using data mining techniques such as classification rules and maximum-relevance-minimum-redundancy. Then, adaptive operator selection algorithms are used to learn which combination of operators works best for a given problem and a given point during the search. This method has been shown to be robust to "poor" heuristics and may lead to saving many (expensive) function evaluations.