Learning and Decision Optimization in Data-Driven Autonomous Systems

Open Access
Author:
Jha, Devesh Kumar
Graduate Program:
Mechanical Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
August 19, 2016
Committee Members:
  • Asok Ray, Dissertation Advisor
  • Asok Ray, Committee Chair
  • Christopher Rahn, Committee Member
  • Thomas Wettergren, Committee Member
  • Minghui Zhu, Outside Member
  • Jan Severin Reimann, Outside Member
  • Abhishek Srivastav, Special Member
Keywords:
  • Autonomous Systems
  • Machine Learning
  • Markov Modeling
  • Combustion Instability
  • Decision Making
  • Motion Planning
Abstract:
Modern human-engineered systems like self-driving cars, smart buildings, power grids etc., are becoming increasingly complex so that they present unparalleled challenges for their design, verification and control. Furthermore, these systems generate data at an unprecedented rate. Even though the data is supposed to be of great help in understanding these systems, it has already presented unique challenges for storage, leave alone their analysis and interpretation (for example, Google is supposed to analyze around 20 Petabytes of data everyday and it is expected that the data generated by systems industry will very soon surpass that). Machine learning provides very powerful statistical techniques for analysis and interpretation of large amounts of data and recently has seen great improvements pushed by large-scale applications in software industries. While there has been huge advancements in the individual fields of machine learning and control theory, unfortunately the modern decision systems for complex systems do not take into account the recent advances of machine intelligence for intelligent control. The role of machine learning in systems engineering is still unexplored; however, it holds tremendous potential for achieving human-like (or better) performance in modern autonomous systems, which is the ultimate goal of Artificial Intelligence. Establishing a synergistic relationship between state-of-the-art control and machine learning techniques holds the key to the synthesis of complex systems with high degrees of autonomy. This thesis explores various concepts and makes contributions to multiple areas for understanding of complex data-driven systems for their decision and control. Among several challenges of complex data-driven systems, one of the central challenge is understanding and representation of data using tools from mathematics and statistics. In this regard, this thesis presents several results for modeling of sequential data using Markov models and Deep Learning. The proposed ideas are tested on various electro-mechanical systems like combustion in lean pre-mixed systems, fatigue in polycrystalline alloys, event detection in sensor networks and lead-acid battery systems. This thesis brings together a completely interdisciplinary approach using techniques from mathematical analysis, theoretical computer science, machine learning, game theory and control theory to present a framework to solve various challenges posed by modern autonomous systems with fundamental contributions to robotic motion planning and temporal data analysis. In order to bring together control theory and machine learning for a class of modern autonomous systems (e.g., automatic driving in urban scenarios with traffic rules), the thesis presents research for the conditions under which we can bring together these fields in an optimal fashion while keeping desired properties like safety, stability, etc.. This thesis makes fundamental contributions to analysis and representation of time-series data using symbolic dynamics and information theory. In particular, algorithms for order estimation and compact representation of Markov models of time-series data are proposed which have computational advantages. Furthermore for analysis of high-dimensional temporal data, an algorithm using Deep learning and Gaussian processes is proposed. The thesis also makes fundamental contributions to motion planning of robotic systems where a class of sampling-based feedback motion planning algorithms are proposed. The algorithm is extended for motion planning of multiple robots using game theory. Furthermore, data-driven algorithms are proposed to bring together learning and planning in complex robotic systems.