Unscented sampling techniques for evolutionary computation with applications to astrodynamic optimization

Abstract

This dissertation investigates several innovative approaches to evolutionary optimization that are relevant to numerous applications in astronautical engineering. The challenges and shortfalls associated with evolutionary algorithms are translated into three overarching goals that directly motivate the research and innovations of this dissertation. The first goal is to investigate and employ techniques that enable evolutionary algorithms to effectively handle constraints in a way that allows for feasible solutions to constrained optimization problems. The second goal is to improve computation times and efficiencies associated with evolutionary algorithms. The last goal is to enhance the evolutionary algorithm's robustness and ability to consistently find accurate solutions within a finite number of iterations. Novel techniques involving the application of unscented sampling, parallel computation, and various forms of exact penalty functions are developed and applied to both genetic algorithms and evolution strategies to achieve these goals. The results of this research offer a promising new set of modified evolutionary algorithms that outperform state-of-the-art techniques on a number of challenging multimodal optimization problems. In addition, these new methods are shown to be very effective in solving a minimum-propellant lunar lander optimal control problem, representing a class of problems that are historically difficult to solve using evolutionary algorithms.