The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Multi-stage stochastic optimization and reinforcement learning for forestry epidemic and covid-19 control planning
Author:	Bushaj, Sabah
View Online:	njit-etd2021-040 (xx, 228 pages ~ 3.8 MB pdf)
Department:	Department of Mechanical and Industrial Engineering
Degree:	Doctor of Philosophy
Program:	Industrial Engineering
Document Type:	Dissertation
Advisory Committee:	Buyuktahtakin-Toy, Ismet Esra (Committee chair) Abdel-Malek, Layek (Committee member) Das, Sanchoy K. (Committee member) Cai, Selina (Committee member) Bilgili, Ecevit Atalay (Committee member)
Date:	2021-08
Keywords:	Epidemic disease modeling Invasive species Mathematical modeling Operations research Reinforcement learning Stochastic programming
Availability:	Unrestricted
Abstract:	This dissertation focuses on developing new modeling and solution approaches based on multi-stage stochastic programming and reinforcement learning for tackling biological invasions in forests and human populations. Emerald Ash Borer (EAB) is the nemesis of ash trees. This research introduces a multi-stage stochastic mixed-integer programming model to assist forest agencies in managing emerald ash borer insects throughout the U.S. and maximize the public benets of preserving healthy ash trees. This work is then extended to present the first risk-averse multi-stage stochastic mixed-integer program in the invasive species management literature to account for extreme events. Significant computational achievements are obtained using a scenario dominance decomposition and cutting plane algorithm.The results of this work provide crucial insights and decision strategies for optimal resource allocation among surveillance, treatment, and removal of ash trees, leading to a better and healthier environment for future generations. This dissertation also addresses the computational difficulty of solving one of the most difficult classes of combinatorial optimization problems, the Multi-Dimensional Knapsack Problem (MKP). A novel 2-Dimensional (2D) deep reinforcement learning (DRL) framework is developed to represent and solve combinatorial optimization problems focusing on MKP. The DRL framework trains different agents for making sequential decisions and finding the optimal solution while still satisfying the resource constraints of the problem. To our knowledge, this is the first DRL model of its kind where a 2D environment is formulated, and an element of the DRL solution matrix represents an item of the MKP. Our DRL framework shows that it can solve medium-sized and large-sized instances at least 45 and 10 times faster in CPU solution time, respectively, with a maximum solution gap of 0.28% compared to the solution performance of CPLEX. Applying this methodology, yet another recent epidemic problem is tackled, that of COVID-19. This research investigates a reinforcement learning approach tailored with an agent-based simulation model to simulate the disease growth and optimize decision-making during an epidemic. This framework is validated using the COVID-19 data from the Center for Disease Control and Prevention (CDC). Research results provide important insights into government response to COVID-19 and vaccination strategies.