Learning and decision-making in spatiotemporally varying domains

Modeling and monitoring of large-scale stochastic phenomena with both spatial and temporal (spatiotemporal) evolution using sensors that can only partially observe the system is a fundamental problem in many applications. In particular, in order to make intelligent decisions in robot coordination and actuation, the robot(s) need to be able to predict the full state of their evolving environment at any given time, given only sparse measurements. The purpose of this dissertation is to take steps towards addressing this challenging problem. The rapid advances in the computational power of compact systems and robotics as a whole has led to an explosion of real-world applications for such distributed cyber-physical systems. This work will present advances in this domain, both in theory and in experiments. It will begin with a description of the Kernel Observers (KO) approach and the derivation of the Evolving Gaussian Processes (E-GP) model for generalizing across similar spatiotemporally evolving systems. It will be shown that method is effective for modeling nonlinear fluid flows, and that there are indeed deep theoretical connections with Koopman operator theory and Dynamic Mode Decomposition (DMD) which have gained recognition in the Computational Flow Dynamics (CFD) community. Further work will be presented on using decomposition analysis of the dynamic model to identify optimal locations for sensors in order to obtain the quickest and most accurate estimate the true state of the system. Next, this dissertation will address a specific distributed cyber-physical systems problem: coordinated multi-agent weeding of a field of crops under conditions of partial environmental information. In conjunction with developing and adapting the E-GP model for this application, this chapter will address the classic explore/exploit problem of making intelligent agent-allocation decisions in a dynamic environment with limited information. The theory developed in these chapters will help inform the final chapter, in which another challenging application is analyzed and solved: visual search in a large-scale environment using a pan-tilt-zoom (PTZ) camera.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Whitman, Joshua Earl

Contributors dc:contributor

• Chowdhary, Girish
• Mehta, Prashant
• Sreenivas, R.S.
• Gazzola, Mattia

• Machine Learning
• Dynamics
• Gaussian Process
• Computational Flow Dynamics
• Traveling Salesman
• Machine Vision
• Koopman Operator
• Agricultural Robotics
• Decision And Control

Statement dc:rights

• Copyright 2022 by Joshua Earl Whitman. All rights reserved.

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/120097

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F120097