Investigating the use of neural network surrogate models in the evolutionary optimization of interplanetary low-thrust trajectories

Abstract

Building on recent advances in the fields of low-thrust trajectory optimization based on shaping methods, Artificial Neural Networks, and surrogate models in Evolutionary Algorithms, an investigation into a novel optimization routine is conducted. A flexible Python tool to evaluate linked trajectories in a two-body model based on the hodographic shaping method is implemented and used to develop an evolutionary optimization approach where a Genetic Algorithm is assisted in finding new candidate solutions by a surrogate model. This surrogate is constructed from previous fitness function evaluations using Machine Learning, specifically by training an Artificial Neural Network. After deriving suitable (hyper-)parameters for the Genetic Algorithm and the Artificial Neural Network an experimental investigation into the algorithm's performance is conducted with a focus on the design of the surrogate for low-thrust trajectory problems. Two example problems based on the Dawn trajectory and the GTOC2 problem are studied. The surrogate approach is able to find good new candidate solutions, i.e. solutions that improve the population's overall fitness, especially when the surrogate is designed to approximate the shaping computation. Additionally, the use of a surrogate pretrained on a general data set of low-thrust transfers is tested and found to considerably improve the initial quality of the model, meaning that more good candidate solutions are found early on, accelerating the algorithm's convergence.