No description available.

No description available.

We introduce two online models for the vertex k-center and the vertex k-median problems. Clients (i.e., graph vertices) and their corresponding links (i.e., graph edges) are revealed sequentially, determining the topology of a graph over time. Clients are revealed by an adversary to an online algorithm that selects existing graph vertices on which to open facilities; once open, a facility cannot be removed or relocated. We define two models: an online algorithm may be restricted to open a facility only at the location of the most recent client or at the location of any existing client. We examine these models on three classes of graphs under two types of adversaries. We establish lower bounds on the respective competitive ratios attainable by any online algorithm for each combination of model, adversary, and graph class. Finally, we describe algorithms whose competitive ratios provide corresponding upper bounds on the best competitive ratios achievable.

In this thesis, we generalize a set of facility location models within a two-stage robust optimization framework by assuming each demand is only known to lie within a continuous and bounded uncertainty region. Our approach involves discretizing each uncertainty region into a set of finite scenarios, each of which represents a potential location where the demand may be realized. We show that the gap between the optimal values of the theorized continuous uncertainty problem and our discretized model can be bounded by a function of the granularity of the discretization. We then propose a solution technique based on row-and-column generation, and compare its performance with existing solution methods. Lastly, we apply our robust location models to the problem of ambulance positioning using cardiac arrest location data from the City of Toronto, and show that hedging against demand location uncertainty may help decrease EMS response times to cardiac arrest emergencies.

In this thesis, we generalize a set of facility location models within a two-stage robust optimization framework by assuming each demand is only known to lie within a continuous and bounded uncertainty region. Our approach involves discretizing each uncertainty region into a set of finite scenarios, each of which represents a potential location where the demand may be realized. We show that the gap between the optimal values of the theorized continuous uncertainty problem and our discretized model can be bounded by a function of the granularity of the discretization. We then propose a solution technique based on row-and-column generation, and compare its performance with existing solution methods. Lastly, we apply our robust location models to the problem of ambulance positioning using cardiac arrest location data from the City of Toronto, and show that hedging against demand location uncertainty may help decrease EMS response times to cardiac arrest emergencies.

No description available.

No description available.

We introduce two online models for the vertex k-center and the vertex k-median problems. Clients (i.e., graph vertices) and their corresponding links (i.e., graph edges) are revealed sequentially, determining the topology of a graph over time. Clients are revealed by an adversary to an online algorithm that selects existing graph vertices on which to open facilities; once open, a facility cannot be removed or relocated. We define two models: an online algorithm may be restricted to open a facility only at the location of the most recent client or at the location of any existing client. We examine these models on three classes of graphs under two types of adversaries. We establish lower bounds on the respective competitive ratios attainable by any online algorithm for each combination of model, adversary, and graph class. Finally, we describe algorithms whose competitive ratios provide corresponding upper bounds on the best competitive ratios achievable.

The views expressed in this thesis are those of the author and do not reflect the official policy or position of the Department of Defense or the U.S. Government. / Determining the best base stationing for military units can be modeled as a capacitated facility location problem with sole sourcing and multiple resource categories. Computational experience suggests that cross decomposition, a unification of Benders Decomposition and Lagrangean relaxation, is superior to other contemporary methods for solving capacitated facility location problems. Recent research extends cross decomposition to pure integer prograrnming problems with explicit application to capacitated facility location problems with sole sourcing; however, this research offers no computational experience. This thesis implements two cross decomposition algorithms for the capacitated facility location problem with sole sourcing and compares these decomposition algorithms with branch and bound methods. For some problems tested, cross decomposition obtains better solutions in less time; however, cross decomposition does not always perform better man branch and bound due to the time required to obtain the cross decomposition bound that is theoretically superior to other decomposition bounds.

The capacitated facility location problem is a well known problem in combinatorial optimization and operations research. In it, we are given a set of clients and a set of possible facility locations. Each client has a certain demand that needs to be satisfied from open facilities, without exceeding their capacity. Whenever we open a facility we incur in a corresponding opening cost. Whenever demand is served, we incur in an assignment cost; depending on the distance the demand travels. The goal is to open a set of facilities that satisfy all demands while minimizing the total opening and assignment costs. In this thesis, we present two novel LP-based approximation algorithms for the capacitated facility location problem. The first algorithm is based on LP-rounding techniques, and is designed for the special case of the capacitated facility location problem where capacities are uniform and assignment costs are given by a tree metric. The second algorithm follows a primal-dual approach, and works for the general case. For both algorithms, we obtain an approximation guarantee that is linear on the size of the problem. To the best of our knowledge, there are no LP-based algorithms known, for the type of instances that we focus on, that achieve a better performance.