(Georgia Institute of Technology, 2023-04-27)
Gandhi, Manan
From aggressive driving, to drone racing, to cardiovascular control, each of these problems involves uncertain dynamics and a requirement to make decisions. There exist both classical and modern techniques for control of dynamical systems which can be used to favor safety, exploration, control performance, or other criteria. Information Theoretic Model Predictive Control (IF-MPC) has been utilized to great effect for autonomous racing by optimizing the system free energy through forward sampling. This dissertation explores the ideas of incorporating safety directly into IF-MPC for uncertain systems. The underpinnings of Model Predictive Path Integral control are presented with its connections to both Information Theory and Bayesian inference. Current research on model learning for biological systems are presented with aim to use more advanced control techniques. The core of the thesis is the development on a framework to perform safe model predictive control of stochastic dynamical systems while taking advantage of traditional control techniques to make decisions during operation. In the development of this novel framework, the foundation of modern, safe control are presented, along with state-of-the-art developments to enable safe control with fewer theoretical and computational restrictions. This work builds upon the idea of augmented importance sampling in IF-MPC to perform tasks under safety constraints for a general system. The control architecture presented here can handle both soft and hard constraints on the state of a dynamical system, as well as balance task completion with the notion of safety.
(Georgia Institute of Technology, 2023-01-13)
Karagoz, Esma
Design and development of complex engineered systems in the aerospace industry have been facing challenges in terms of managing ever-increasing complexity. Due to this complexity, engineering design problems become ill-defined by nature; in other words, as the design problem is gradually solved, it becomes better understood with formal specifications. Hence, designing an engineered system is iterative and occurs in multiple stages. In addition, as the complexity of the system increases, so does the need for extensive collaboration between different disciplinary teams to manage the interdependencies between system elements. However, due to the complexity and integration challenges in current multidisciplinary design efforts, the behavior of the system may not be accurately predicted, and the project outcomes may not be aligned with what designers have in mind. Hence, better-informed decisions need to be made during the design and development of complex systems. This is one of the most prevalent challenges observed in the aerospace industry, especially in the integration phase of product development, causing budget and schedule overruns.
In order to keep up with rapidly changing customer expectations in a timely manner and within a given budget, better-informed decisions must be made by utilizing innovative engineering methods. There is an increasing demand for the integration and interoperability of multiple intelligent systems to automate engineering processes and minimize human involvement. This allows for autonomous operation and intelligent decision-making throughout the system life cycle. In light of this, this work develops a methodology to automate decision-making processes by modeling various types of data and knowledge in a systems engineering environment. In order to combine this heterogeneous data and knowledge, the following elements are required: (1) semantic modeling through the model-based system and architecture descriptions in MBSE environments, (2) physics-based modeling that captures the details of multidisciplinary design and optimization analyses, and (3) mathematical modeling to describe computational decision-making algorithms. By combining this heterogeneous knowledge, a custom ontological metamodel is created to describe the connections between the semantic and analytical models that also allow for communication with external engineering environments. This metamodel is used as a guideline to build a system model for the case study. In order to build interoperability between the system model and the external engineering tools, the data in the system model is exported and converted into property graphs for better data storage, and for enabling machine learning tools to use this data as input. These property graphs represent the knowledge base for the decision-making algorithm. Having built a knowledge base, heterogeneous graph neural networks are used to assess whether there is missing information in the knowledge base. The reason is that ontologies are incomplete by nature as they are built iteratively and collaboratively, making them prone to lacking important information. After securing a complete knowledge base, graph algorithms are used for knowledge graph reasoning, which provides insights and recommendations along with the explanation paths behind those, to aid systems engineers when making decisions.
Overall, the expected contributions of this research to the aerospace industry are (1) aiding systems engineers when making design decisions, (2) enabling systems engineers to capture previous knowledge and experiences, (3) improving early identification of unexpected failures causing additional efforts, (4) efficient training of novice engineers through the established knowledge base, and (5) reducing development times and budget.