Event
Ph.D. Dissertation Defense: Eduardo Romero Arvelo
Wednesday, November 9, 2016
10:00 a.m.
2328 AVW Bldg.
Maria Hoo
301 405 3681
mch@umd.edu
Advisory Committee:
Professor Nuno Martins, Chair/Advisor
Professor Gilmer Blankenship
Professor Pamela Abshire
Professor Timothy Horiuchi
Professor Nikhil Chopra, Dean's representative
Date/Time: Wednesday, November 9, 2016 at 10am
Title: Maximal Persistent Surveillance and Optimal Sensor Usage in Denied Environments
Abstract:
In the first part of the thesis, we design control policies for Markov decision processes (MDP) with the objective of generating the maximal set of recurrent states, subject to convex constraints on the set of invariant probability mass functions. We propose a design method for memoryless policies and fully observable MDPs with finite state and action spaces. Our approach relies on a finitely parametrized convex program inspired by principles of entropy maximization.
Next, we explore the problem of designing controllers for autonomous robots tasked with maximal persistent surveillance of an area in which there are forbidden regions. We model each robot as an MDP whose state comprises its position on a finite two-dimensional lattice and the direction of motion. The goal is to find the minimum number of robots and an associated time-invariant memoryless control policy that guarantees that the largest number of states are persistently surveilled without ever visiting a forbidden state.
In the second part of the thesis, we study the problem of optimal sensor usage in denied environments, inside which state observations are only available by incurring an extra cost. Observations outside the denied environment are cost free. The goal is to understand the trade-off between paying to access the sensor immediately, and waiting for a free sensor use should the system exit the denied environment. We show that the analysis of this problem simplifies by recasting it as renewal reward process, which enables us to establish conditions for which the cost has a unique minimum, if it exists, thus facilitating the search for its minimizer.
Finally, we extend these results to the case when the state space is discrete and the state update is ruled by a Markov chain. We establish conditions on the initial distribution of the Markov chain that guarantee that the cost has a unique minimum, if it exists. In particular, these conditions rely on constraining the radial stochastic order of an auxiliary Markov process. By analyzing these problems, we hope to provide valuable insights into the core of some of the challenges that may arise in real-world implementations