Recent developments in mixed integer programming have highlighted the need for multi-row cuts. To this day, the performance of such cuts has typically fallen short of the single-row Gomory mixed integer cut. This disparity between the theoretical need and the practical shortcomings of multi-row cuts motivates the study of both the mixed integer cut and multi-row cuts. In this thesis, we build on the theoretical foundations of the mixed integer cut and develop techniques to derive multi-row cuts. The first chapter introduces the mixed integer programming problem. In this chapter, we review the terminology and cover some basic results that find application throughout this thesis. Furthermore, we describe the practical solution of mixed integer programs, and in particular, we discuss the role of cutting planes and our contributions to this theory. In Chapter 2, we investigate the Gomory mixed integer cut from the perspective of group polyhedra. In this setting, the mixed integer cut appears as a facet of the master cyclic group polyhedron. Our chief contribution is a characterization of the adjacent facets and the extreme points of the mixed integer cut. This provides insight into the families of cuts that may work well in conjunction with the mixed integer cut. We further provide extensions of these results under mappings between group polyhedra. For the remainder of this thesis we explore a framework for deriving multi-row cuts. For this purpose, we favor the method of superadditive lifting. This technique is largely driven by our ability to construct superadditive under-approximations of a special value function known as the lifting function. We devote our effort to precisely this task. Chapter 3 reviews the theory behind superadditive lifting and returns to the classical problem of lifted flow cover inequalities. For this specific example, the lifting function we wish to approximate is quite complicated. We overcome this difficulty by adopting an indirect method for proving the validity of a superadditive approximation. Finally, we adapt the idea to high-dimensional lifting problems, where evaluating the exact lifting function often poses an immense challenge. Thus we open entirely unexplored problems to the powerful technique of lifting. Next, in Chapter 4, we consider the computational aspects of constructing strong superadditive approximations. Our primary contribution is a finite algorithm that constructs non-dominated superadditive approximations. This can be used to build superadditive approximations on-the-fly to strengthen cuts derived during computation. Alternately, it can be used offline to guide the search for strong superadditive approximations through numerical examples. We follow up in Chapter 5 by applying the ideas of Chapters 3 and 4 to high-dimensional lifting problems. By working out explicit examples, we are able to identify non-dominated superadditive approximations for high-dimensional lifting functions. These approximations strengthen existing families of cuts obtained from single-row relaxations. Lastly, we show via the stable set problem how the derivation of the lifting function and its superadditive approximation can be entirely embedded in the computation of cuts. Finally, we conclude by identifying future avenues of research that arise as natural extensions of the work in this thesis.

This research shows optimization approaches to protein folding. The protein folding problem is to predict the compact three dimensional structure of a protein based on its amino acid sequence. This research focuses on ab-initio mathematical models to find provably optimal solutions to the 2D HP-lattice protein folding model. We built two integer programming (IP) models and five constraint programming (CP) models. All the models give provably optimal solutions. We also developed some CP techniques to solve the problem faster and then compared their computational times. We tested the models with several protein instances. My models, while they are probably too slow to use in practice, are significantly faster than the alternatives, and thus are mathematically relevant. We also provided reasons why protein folding is hard using complexity analysis. This research will contribute to showing whether CP can be an alternative to or a complement of IP in the future. Moreover, figuring out techniques combining CP and IP is a prominent research issue and our work will contribute to that literature. It also shows which IP/CP strategies can speed up the running time for this type of problem. Finally, it shows why a mathematical approach to protein folding is especially hard not only mathematically, i.e. NP-hard, but also practically.

Many supply chain problems involve discrete decisions in a dynamic environment. The inventory routing problem is an example that combines the dynamic control of inventory at various facilities in a supply chain with the discrete routing decisions of a fleet of vehicles that moves product between the facilities. We study these problems modeled as mixed-integer programs and propose a time decomposition based on approximate inventory valuation. We generate the approximate value function with an algorithm that combines data fitting, discrete optimization and dynamic programming methodology. Our framework allows the user to specify a class of piecewise linear, concave functions from which the algorithm chooses the value function. The use of piecewise linear concave functions is motivated by intuition, theory and practice. Intuitively, concavity reflects the notion that inventory is marginally more valuable the closer one is to a stock-out. Theoretically, piecewise linear concave functions have certain structural properties that also hold for finite mixed-integer program value functions. (Whether the same properties hold in the infinite case is an open question, to our knowledge.) Practically, piecewise linear concave functions are easily embedded in the objective function of a maximization mixed-integer or linear program, with only a few additional auxiliary continuous variables. We evaluate the solutions generated by our value functions in a case study using maritime inventory routing instances inspired by the petrochemical industry. The thesis also includes two other contributions. First, we review various data fitting optimization models related to piecewise linear concave functions, and introduce new mixed-integer programming formulations for some cases. The formulations may be of independent interest, with applications in engineering, mixed-integer non-linear programming, and other areas. Second, we study a discounted, infinite-horizon version of the canonical single-item lot-sizing problem and characterize its value function, proving that it inherits all properties of interest from its finite counterpart. We then compare its optimal policies to our algorithm's solutions as a proof of concept.

In this thesis, we use integer programming techniques to solve the resource constrained shortest path problem (RCSPP) which seeks a minimum cost path between two nodes in a directed graph subject to a finite set of resource constraints. Although NP-hard, the RCSPP is extremely useful in practice and often appears as a subproblem in many decomposition schemes for difficult optimization problems. We begin with a study of the RCSPP polytope for the single resource case and obtain several new valid inequality classes. Separation routines are provided, along with a polynomial time algorithm for constructing an auxiliary conflict graph which can be used to separate well known valid inequalities for the node packing polytope. We establish some facet defining conditions when the underlying graph is acyclic and develop a polynomial time sequential lifting algorithm which can be used to strengthen one of the inequality classes. Next, we outline a branch-and-cut algorithm for the RCSPP. We present preprocessing techniques and branching schemes which lead to strengthened linear programming relaxations and balanced search trees, and the majority of the new inequality classes are generalized to consider multiple resources. We describe a primal heuristic scheme that uses fractional solutions, along with the current incumbent, to search for new feasible solutions throughout the branch-and-bound tree. A computational study is conducted to evaluate several implementation choices, and the results demonstrate that our algorithm outperforms the default branch-and-cut algorithm of a leading integer programming software package. Finally, we consider the dial-a-flight problem (DAFP), a new vehicle routing problem that arises in the context of on-demand air transportation and is concerned with the scheduling of a set of travel requests for a single day of operations. The DAFP can be formulated as an integer multicommodity network flow model consisting of several RCSPPs linked together by set partitioning constraints which guarantee that all travel requests are satisfied. Therefore, we extend our branch-and-cut algorithm for the RCSPP to solve the DAFP. Computational experiments with practical instances provided by the DayJet Corporation verify that the extended algorithm also outperforms the default branch-and-cut algorithm of a leading integer programming software package.

In this thesis we develop integer programming approaches for solving network flow problems with equal-split restrictions. Such problems arise in traffic engineering of internet protocol networks. Equal-split structure is used in protocols like OSPF and IS-IS that allow flow to be split among the multiple shortest paths. Equal-split assumptions also arise in peer-to-peer networks and road optimization problems. All the previous work on this problem has been focused on developing heuristic methods for the specific applications. We are the first ones to study the problem as a general network flow problem and provide a polyhedral study. First we consider a general multi-commodity network flow problem with equal split restrictions. This problem is NP-hard in general. We perform a polyhedral study on mixed integer linear programming formulation for this problem. Valid inequalities are obtained, and are incorporated within a branch-and-cut framework to solve the problem. We provide fast separation schemes for most of the families of valid inequalities. Computational results are presented to show the effectiveness of cutting plane families. Next, we consider the OSPF weight setting problem. We propose an integer programming formulation for this problem. A decomposition based approach to solve the problem is presented next. Valid inequalities, exploiting the structure, are obtained for this problem. We also propose heuristic methods to get good starting solutions for the problem. The proposed cutting planes and heuristic methods are integrated within a branch-and-cut framework to solve the problem. We present computational experiments that demonstrate the effectiveness of our approach to obtain solutions with tight optimality gaps as compared with default CPLEX. Finally, we consider an equal split flow problem on bipartite graphs. We present an integer programming formulation for this problem that models the equal-split in a different way than the multi-commodity network flow problem discussed before. Valid inequalities and heuristic methods for this problem are proposed, and are integrated within the branch-and-cut framework. We present computational experiments demonstrating the effectiveness of our solution strategy. We present an alternate formulation for the problem with some favorable polyhedral properties. Lastly, a computational comparison between the two formulations is presented.

Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, 1981 / Vita. / Bibliography: leaves 97-98. / by Dorothy Elliott Sempolinski. / Ph. D. / Ph. D. Massachusetts Institute of Technology, Sloan School of Management

Stochastic programming is an optimization technique that incorporates random variables as parameters. Because it better reflects the uncertain real world than its traditional deterministic counterpart, stochastic programming has drawn increasingly more attention among decision-makers, and its applications span many fields including financial engineering, health care, communication systems, and supply chain management. On the flip side, stochastic programs are usually very difficult to solve, which is further compounded by the fact that in many of the aforementioned applications, we also have discrete decisions, thereby rendering these problems even more challenging. In this dissertation, we study the class of two-stage stochastic mixed-integer programs (SMIP), which, as its name suggests, lies at the confluence of two formidable classes of problems. We design a novel algorithm for this class of problems, and also explore specialized approaches for two related real-world applications. Although a number of algorithms have been developed to solve two-stage SMIPs, most of them deal with problems containing purely integer or continuous variables in either or both of the two stages, and frequently require the technology and/or recourse matrices to be deterministic. As a ground-breaking effort, in this work, we address the challenging class of two-stage SMIPs that involve 0-1 mixed-integer variables in both stages. The only earlier work on solving such problems (Carøe and Schultz (1999)) requires the optimization of several non-smooth Lagrangian dual problems using subgradient methods in the bounding process, which turns out to be computationally very expensive. We begin with proposing a decomposition-based branch-and-bound (DBAB) algorithm for solving two-stage stochastic programs having 0-1 mixed-integer variables in both stages. Since the second-stage problems contain binary variables, their value functions are in general nonconvex and discontinuous; hence, the classical Benders' decomposition approach (or the L-shaped method) for solving two-stage stochastic programs, which requires convex subproblem value functions, cannot be directly applied. This motivates us to relax the second-stage problems and accompany this relaxation with a convexification process. To make this process computationally efficient, we propose to construct a certain partial convex hull representation of the two-stage solution space, using the relaxed second-stage constraints and the restrictions confining the first-stage variables to lie within some hyperrectangle. This partial convex hull is sequentially generated using a convexification scheme, such as the Reformulation-Linearization Technique (RLT), which yields valid inequalities that are functions of the first-stage variables and, of noteworthy importance, are reusable in the subsequent subproblems by updating the values of the first-stage variables. Meanwhile, since the first stage contains continuous variables, whenever we tentatively fix these variables at some given feasible values, the resulting constraints may not be facial with respect to the associated bounding constraints that are used to construct the partial convex hull. As a result, the constructed Benders' subproblems define lower bounds for the second-stage value functions, and likewise, the resulting Benders' master problem provides a lower bound for the original stochastic program defined over the same hyperrectangle. Another difficulty resulting from continuous first-stage variables is that when the given first-stage solution is not extremal with respect to its bounds, the second-stage solution obtained for a Benders' subproblem defined with respect to a partial convex hull representation in the two-stage space may not satisfy the model's binary restrictions. We thus need to be able to detect whether or not a Benders' subproblem is solved by a given fractional second-stage solution. We design a novel procedure to check this situation in the overall algorithmic scheme. A key property established, which ensures global convergence, is that these lower bounds become exact if the given first-stage solution is a vertex of the defining hyperrectangle, or if the second-stage solution satisfies the binary restrictions. Based on these algorithmic constructs, we design a branch-and-bound procedure where the branching process performs a hyperrectangular partitioning of the projected space of the first-stage variables, and lower bounds for the nodal problems are computed by applying the proposed modified Benders' decomposition method. We prove that, when using the least-lower-bound node-selection rule, this algorithm converges to a global optimal solution. We also show that the derived RLT cuts are not only reusable in subsequent Benders iterations at the same node, but are also inheritable by the subproblems of the children nodes. Likewise, the Benders' cuts derived for a given sub-hyperrectangle can also be inherited by the lower bounding master programs solved for its children nodes. Using these cut inheritance properties results in significant savings in the overall computational effort. Some numerical examples and computational results are presented to demonstrate the efficacy of this approach. The sizes of the deterministic equivalent of our test problems range from having 386 continuous variables, 386 binary variables, and 386 constraints, up to 1795 continuous variables, 1539 binary variables, and 1028 constraints. The results reveal an average savings in computational effort by a factor of 9.5 in comparison with using a commercial mixed-integer programming package (CPLEX 8.1) on a deterministic equivalent formulation. We then explore an important application of SMIP to enhance the traditional airline fleet assignment models (FAM). Given a flight schedule network, the fleet assignment problem solved by airline companies is concerned with assigning aircraft to flight legs in order to maximize profit with respect to captured path- or itinerary-based demand. Because certain related crew scheduling regulations require early information regarding the type of aircraft serving each flight leg, the current practice adopted by airlines is to solve the fleet assignment problem using estimated demand data 10-12 weeks in advance of departure. Given the level of uncertainty, deterministic models at this early stage are inadequate to obtain a good match of aircraft capacity with passenger demands, and revisions to the initial fleet assignment become naturally pertinent when the observed demand differs considerably from the assigned aircraft capacities. From this viewpoint, the initial decision should embrace various market scenarios so that it incorporates a sufficient look-ahead feature and provides sufficient flexibility for the subsequent re-fleeting processes to accommodate the inevitable demand fluctuations. With this motivation, we propose a two-stage stochastic programming approach in which the first stage is concerned with the initial fleet assignment decisions and, unlike the traditional deterministic methodology, focuses on making only a family-level assignment to each flight leg. The second stage subsequently performs the detailed assignments of fleet types within the allotted family to each leg under each of the multiple potential scenarios that address corresponding path- or itinerary-based demands. In this fashion, the initial decision of what aircraft family should serve each flight leg accomplishes the purpose of facilitating the necessary crew scheduling decisions, while judiciously examining the outcome of future re-fleeting actions based on different possible demand scenarios. Hence, when the actual re-fleeting process is enacted several weeks later, this anticipatory initial family-level assignment will hopefully provide an improved overall fleet type re-allocation that better matches demand. This two-stage stochastic model is complemented with a secondary model that performs adjustments within each family, if necessary, to provide a consistent fleet type-assignment information for accompanying decision processes, such as yield management. We also propose several enhanced fleet assignment models, including a robust optimization model that controls decision variation among scenarios and a stochastic programming model that considers the recapture effect of spilled demand. In addition to the above modeling concepts and framework, we also contribute in developing effective solution approaches for the proposed model, which is a large-scale two-stage stochastic 0-1 mixed-integer program. Because the most pertinent information needed from the initial fleet assignment is at the family level, and the type-level assignment is subject to change at the re-fleeting stage according to future demand realizations, our solution approach focuses on assigning aircraft families to the different legs in the flight network at the first stage, while finding relaxed second-stage solutions under different demand scenarios. Based on a polyhedral study of a subsystem extracted from the original model, we derive certain higher-dimensional convex hull as well as partial convex hull representations for this subsystem. Accordingly, we propose two variants for the primary model, both of which relax the binary restrictions on the second-stage variables, but where the second variant then also accommodates the partial convex hull representations, yielding a tighter, albeit larger, relaxation. For each variant, we design a suitable solution approach predicated on Benders' decomposition methodology. Using certain realistic large-scale flight network test problems having 900 flight legs and 1,814 paths, as obtained from United Airlines, the proposed stochastic modeling approach was demonstrated to increase daily expected profits by about 3% (which translates to about $160 million per year) in comparison with the traditional deterministic model in present usage, which considers only the expected demand. Only 1.6% of the second-stage binary variables turn out to be fractional in the first variant, and this number is further reduced to 1.2% by using the tighter variant. Furthermore, when attempting to solve the deterministic equivalent formulation for these two variants using a commercial mixed-integer programming package (CPLEX 8.1), both the corresponding runs were terminated after reaching a 25-hour cpu time limit. At termination, the software was still processing the initial LP relaxation at the root node for each of these runs, and no feasible basis was found. Using the proposed algorithms, on the other hand, the solution times were significantly reduced to 5 and 19 hours for the two variants, respectively. Considering that the fleet assignment models are solved around three months in advance of departure, this solution time is well acceptable at this early planning stage, and the improved quality in the solution produced by considering the stochasticity in the system is indeed highly desirable. Finally, we address another practical workforce planning problem encountered by a global financial firm that seeks to manage multi-category workforce for functional areas located at different service centers, each having office-space and recruitment-capacity constraints. The workforce demand fluctuates over time due to market uncertainty and dynamic project requirements. To hedge against the demand fluctuations and the inherent uncertainty, we propose a two-stage stochastic programming model where the first stage makes personnel recruiting and allocation decisions, while the second stage, based on the given personnel decision and realized workforce demand, decides on the project implementation assignment. The second stage of the proposed model contains binary variables that are used to compute and also limit the number of changes to the original plan. Since these variables are concerned with only one quality aspect of the resulting workforce plan and do not affect feasibility issues, we replace these binary variables with certain conservative policies regarding workforce assignment change restrictions in order to obtain more manageable subproblems that contain purely continuous variables. Numerical experiments reveal that the stochastic programming approach results in significantly fewer alterations to the original workforce plan. When using a commercial linear programming package CPLEX 9.0 to solve the deterministic equivalent form directly, except for a few small-sized problems, this software failed to produce solutions due to memory limitations, while the proposed Benders' decomposition-based solution approach consistently solved all the practical-sized test problems with reasonable effort. To summarize, this dissertation provides a significant advancement in the algorithmic development for solving two-stage stochastic mixed-integer programs having 0-1 mixed-integer variables in both stages, as well as in its application to two important contemporary real-world applications. The framework for the proposed solution approaches is to formulate tighter relaxations via partial convex hull representations and to exploit the resulting structure using suitable decomposition methods. As decision robustness is becoming increasingly relevant from an economic viewpoint, and as computer technological advances provide decision-makers the ability to explore a wide variety of scenarios, we hope that the proposed algorithms will have a notable positive impact on solving stochastic mixed-integer programs. In particular, the proposed stochastic programming airline fleet assignment and the workforce planning approaches studied herein are well-poised to enhance the profitability and robustness of decisions made in the related industries, and we hope that similar improvements are adapted by more industries where decisions need to be made in the light of data that is shrouded by uncertainty. / Ph. D.

Integer programming can be used to provide solutionsto complex decision and planning problems occurring in a wide varietyof situations. Applying integer programming to a real life problembasically involves a first phase where a mathematical model isconstructed, and a second phase where the problem described by themodel is solved. While the nature of the challenges involved in therespective two phases differ, the strong relationship between theproperties of models, and which methods that are appropriate for theirsolution, links the two phases. This thesis constitutes of threepapers, of which the third one considers the modeling phase, while thefirst and second one consider the solution phase.   Many applications of column generation yield master problems of setpartitioning type, and the first and second papers presentmethodologies for solving such problems. The characteristics of themethodologies presented are that all successively found solutions arefeasible and integral, where the retention of integrality is a majordistinction from other column generation methods presented in theliterature.   The third paper concerns nurse scheduling and describes the results ofa pilot implementation of a scheduling tool at a Swedish nursing ward.This paper focuses on the practical aspects of modeling and thechallenges of providing a solution to a complex real life problem.

Internal audit departments strive to control risk within an organization. To do this they choose specific audit areas to include in an audit plan. In order to select areas, they usually focus on those areas with the highest risk. Even though high risk areas are considered, there are various other restrictions such as resource constraints (in terms of funds, manpower and hours) that must also be considered. In some cases, management might also have special requirements. Traditionally this area selection process is conducted using manual processes and requires significant decision maker experience. This makes it difficult to take all possibilities into consideration while also catering for all resource constraints and special management requirements. In this study, mathematical techniques used in capital budgeting problems are explored to solve the IT audit area selection problem. A DSS is developed which implements some of these mathematical techniques such as a linear programming model, greedy heuristic, improved greedy heuristic and evolutionary heuristic. The DSS also implements extensions to the standard capital budgeting model to make provision for special management requirements. The performance of the mathematical techniques in the DSS is tested by applying different decision rules to each of the techniques and comparing those results. The DSS, empirical experiments and results are also presented in this research study. Results have shown that in most cases a binary 0-1 model outperformed the other techniques. Internal audit management should therefore consider this model to assist with the construction of an IT internal audit plan. / MSc (Computer Science), North-West University, Potchefstroom Campus, 2015

Internal audit departments strive to control risk within an organization. To do this they choose specific audit areas to include in an audit plan. In order to select areas, they usually focus on those areas with the highest risk. Even though high risk areas are considered, there are various other restrictions such as resource constraints (in terms of funds, manpower and hours) that must also be considered. In some cases, management might also have special requirements. Traditionally this area selection process is conducted using manual processes and requires significant decision maker experience. This makes it difficult to take all possibilities into consideration while also catering for all resource constraints and special management requirements. In this study, mathematical techniques used in capital budgeting problems are explored to solve the IT audit area selection problem. A DSS is developed which implements some of these mathematical techniques such as a linear programming model, greedy heuristic, improved greedy heuristic and evolutionary heuristic. The DSS also implements extensions to the standard capital budgeting model to make provision for special management requirements. The performance of the mathematical techniques in the DSS is tested by applying different decision rules to each of the techniques and comparing those results. The DSS, empirical experiments and results are also presented in this research study. Results have shown that in most cases a binary 0-1 model outperformed the other techniques. Internal audit management should therefore consider this model to assist with the construction of an IT internal audit plan. / MSc (Computer Science), North-West University, Potchefstroom Campus, 2015