Date of Completion
Bryan Weber; Cara Battersby; Jonathan Trump
University Scholar Major
Artificial Intelligence and Robotics | Astrodynamics | Other Astrophysics and Astronomy
This project aims to determine the feasibility of using NeuroEvolution of Augmenting Topologies (NEAT), an advanced neural network evolution scheme, to optimize orbital transfer trajectories. More specifically, this project compares a genetically evolved neural network to a standard Hohmann transfer between Earth and Mars. To test these two methods, an N-body simulation environment was created to accurately determine the result of gravitational interactions on a theoretical spacecraft when combined with planned engine burns. Once created, this simulation environment was used to train the neural networks created using the NEAT Python module. A genetic algorithm was used to modify the topology of the network in addition to the traditional weight and bias modifications to produce a highly effective individual or batch of individuals that can process various positional and velocity inputs to generate an efficient orbital transfer. The performance of these neural networks was measured by comparing the transfer burn plans they generate to the standard Hohmann transfer using a variety of factors such as transfer time and fuel consumption. This paper presents a background in neural networks, genetic algorithms, and NEAT, discusses the methods chosen for this specific project, and summarizes and draws conclusions from the results of the neural network training.
Ultimately, it was found that the created program was effective in training neural networks to optimize for either time, delta-v, or a combination of both. More specifically, the neural networks consistently created solutions that were more time-efficient than the standard Hohmann transfer, could make equally effective solutions when considering time and delta-v in equal weights, and did not create effective solutions when optimizing for only delta-v.
Wetherell, Nathan, "Optimization of Orbital Trajectories Using NeuroEvolution of Augmenting Topologies" (2022). University Scholar Projects. 81.