Troy Dontigney, United States Air Force; Laurence Merkle, Air Force Institute of Technology; Richard Cobb, Air Force Institute of Technology; John Colombi, Air Force Institute of Technology; Gary Lamonth, Air Force Institute of Technology
Keywords: Space Situational Awareness, Multi-Objective Evolutionary Algorithms, Pareto Front Comparison, High-Performance Computing, Modeling and Simulation
Abstract:
Geosynchronous Earth Orbit (GEO) is a valuable orbital regime containing many mission-critical national and civilian assets. As the region becomes increasingly congested, the risk of collision between resident space objects (RSOs) will grow as well. The need for a space surveillance network (SSN) capable of resolving GEO targets with greater resolution and frequency is imminent. Many research efforts have aimed at developing computationally feasible models and methods for finding the SSN architectures capable of detecting sufficiently small RSOs often enough to maintain an accurate picture of the state of objects in GEO at the lowest possible price. Several recent efforts investigated using high-performance computing (HPC) and evolutionary algorithms (EA) to identify near-optimal solutions in this vast search space. Stern and Wachtel proposed a 28-dimension model of SSNs, coupled with the Non-dominated Sorted Genetic Algorithm II (NSGA-II) as one such method. Their model includes four classes of optical sensors: ground-based telescopes (GBTs), equatorial observation satellites (obsats), sun-synchronous obsats, and near-GEO obsats. Within their model, GBTs may be placed at zero or more user-defined locations on Earth with one or more telescopes at each location. If included in an architecture, there may be only one constellation of each class of obsat, appearing at one of several possible altitudes, with a variable number of obsats in the constellations. Each constellation or GBT site may use different aperture diameters, but must use the same diameter within the given constellation or site. Their methodology can be broadly summarized in four steps: 1) simulate 813 targets using AGIs Systems Toolkit (STK) to approximate a realistic catalog of GEO satellites, 2) compute access, range, angular parameters, and illumination conditions pairwise between each sensor and target for each single-class configuration possible in the model, 3) combine that data to determine performance of any of the more than 2 x 1021 architectures possible in the model, and 4) use NSGA-II to perform a non-exhaustive exploration of the solution space. For this application, the search is orders of magnitude more computationally expensive than the simulations.
Extending that effort, the current research increases efficacy of those methods by presenting a methodology for selecting the most effective from a selection of multi-objective (MO) search algorithms. Four different methods of comparing Pareto fronts are examined and applied to a selection of common MO search algorithms in order to determine the strengths and weaknesses of the methods when applied to this search space. Additionally, a quantitative investigation of the tradeoffs associated with using a subset of the simulation data to reduce the computational load of evaluating multiple algorithms is presented. Given the substantial cost of running a single search, and the projected growth of RSOs in GEO, it is desirable to reduce computational cost while evaluating multiple algorithms. Therefore, subsets of 20, 81, 203, and 407 targets (2.5%, 10%, 25%, and 50% of the full set, respectively) are analyzed statistically, compared with the original 813 RSO set, and used to repeat the comparisons between algorithms. The results are analyzed to demonstrate the tradeoffs involved in using subsets of data to quickly evaluate multiple algorithms, recommending a way forward for analyzing similar large-scale problems of interest to the DoD.
Date of Conference: September 17-20, 2019
Track: Space Situational Awareness