Essential for services such as communication, navigation, and weather prediction, satellite constellations must be designed to minimize loss of coverage while subject to constraints on space traffic and the number of satellites available. Genetic algorithms (GAs) offer a versatile method of optimization with demonstrated success in applied problems. However, in the case of a computationally expensive problem, such as satellite coverage of the earth, the necessity of repeated fitness evaluations prevents convergence of a GA in a feasible time frame. Implementing a more efficient surrogate model to estimate the expensive objective function poses a potential solution to this dilemma. Recent advances in machine learning methods, in particular neural networks, make them a compelling candidate as a surrogate function. A genetic algorithm incorporating an ensemble of neural networks as a surrogate function is evaluated on a set of canonical test problems, including those with discrete inputs and multimodal objective functions, and finally applied to the problem of constellation design.

This research was conducted at the Research in Industrial Projects for Students (RIPS) REU at the Institute for Pure and Applied Mathematics (UCLA) and sponsored by The Aerospace Corporation.


This work was presented both at IPAM and The Aerospace Corporation.