INDIGO Home University of Illinois at Urbana-Champaign logo uic building uic pavilion uic student center

Methods in Large Scale Inverse Optimal Control

Show full item record

Bookmark or cite this item:

Files in this item

File Description Format
PDF MONFORT-DISSERTATION-2016.pdf (1MB) (no description provided) PDF
Title: Methods in Large Scale Inverse Optimal Control
Author(s): Monfort, Mathew
Advisor(s): Ziebart, Brian
Contributor(s): Berger-Wolf, Tanya; Gmytrasiewicz, Piotr; Reyzin, Lev; Carr, Peter; Ziebart, Brian
Department / Program: Computer Science
Degree Granting Institution: University of Illinois at Chicago
Degree: PhD, Doctor of Philosophy
Genre: Doctoral
Subject(s): machine learning artificial intelligence inverse optimal control graph search autonomous agents reinforcement learning path distributions robotic control robotics robots activity recognition
Abstract: As our technology continues to evolve, so does the complexity of the problems that we expect our systems to solve. The challenge is that these problems come at increasing scales that require innovative solutions in order to be tackled efficiently. The key idea behind Inverse Optimal Control (IOC) is that we can learn to emulate how a human completes these complex tasks by modeling the observed decision process. This thesis presents algorithms that extend the state-of-the art in IOC in order to efficiently learn complex models of human behavior. We explore the use of an admissible heuristic in estimating path distributions through weighted graphs. This includes a modified version of the softened policy iteration method used in Maximum Entropy Inverse Optimal Control and present the SoftStar algorithm which merges ideas from Maximum Entropy IOC and A* Search for an efficient probabilistic search method that estimates path distributions through weighted graphs with approximation guarantees. We then explore IOC methods for prediction and planning in problems with linear dynamics that require real-time solutions. This includes an inverse linear quadratic regulation (LQR) method for efficiently predicting intent in 3-dimensional space and a discrete-continuous hybrid version of inverse LQR that uses discrete waypoints to guide the continuous LQR distribution. The presented techniques are evaluated on a number of different problem settings including planning trajectories of handwritten characters, modeling the ball-handler decision process in professional soccer, predicting intent in completing household tasks, and planning robotic motion trajectories through a cluttered workspace.
Issue Date: 2016-11-23
Type: Thesis
Rights Information: Copyright 2017 Monfort, Mathew
Date Available in INDIGO: 2017-02-17
Date Deposited: December 2

This item appears in the following Collection(s)

Show full item record


Country Code Views
United States of America 91
China 16
Russian Federation 14
Ukraine 13
Iran 12


My Account


Access Key