Ph.D. Theses
http://digitalarchive.boun.edu.tr/handle/123456789/11650
Tue, 28 May 2024 00:59:22 GMT2024-05-28T00:59:22ZData-driven local search heuristics for bilevel network design problems
http://digitalarchive.boun.edu.tr/handle/123456789/21454
Data-driven local search heuristics for bilevel network design problems
Sevim, İsmail.
In the Network Design Problem (NDP), one aims to design the configuration of a network by installing links between a set of given nodes and determine the flow of a set of commodities over these installed links. In this thesis, we work on two bilevel NDPs where the sequential process of decision making approach is inherited. In the first bilevel NDP we model the strategic flight NDP of a small airline carrier as a network interdiction problem to analyse the maximum possible disruption in its flight network in the wake of virtual attacks performed by a competitor. We call this problem the r - Interdiction Network Design Problem with Lost Demand (RI-NDPLD). In the second problem, namely Bilevel Optimization Model for the Reconfiguration of refugee camp network (BOpt-RRC), the readjustment of configurations of refugee camp network are studied under the case of new refugee flows and possible variations in the supply of public service providers. We implement a set of generic local search matheuristics to solve both problems. In the Tabu Search (TS) proposed for the RI- NDPLD, we enhance the generic implementation with bound based pruning and regression based candidate solution set generation procedures to reduce the computational burden of explicit evaluation of all neighboring solutions, and hence, enjoy better diversification. We also implement a generic TS to solve the BOpt-RRC and devise an adaptive neighborhood selection procedure to incorporate into this implementation. In addition to the generic TS, we also implement a Variable Neighborhood Search (VNS) matheuristic and devise an association rule based injection procedure to incorporate good solution components to initial solutions obtained by usual random shaking. Experimental studies reveal promising results for the proposed methods.
Sat, 01 Jan 2022 00:00:00 GMThttp://digitalarchive.boun.edu.tr/handle/123456789/214542022-01-01T00:00:00ZDistance-based learning approaches for multiple instance learning
http://digitalarchive.boun.edu.tr/handle/123456789/19800
Distance-based learning approaches for multiple instance learning
Sivrikaya, Özgür Emre.
Multiple Instance Learning (MIL) is a weakly supervised approach that focuses on the labeling of a set of instances (i.e. bags) where the label information of in dividual instances is generally unknown. Many of the earlier MIL studies focus on certain assumptions regarding the relationship between the bag and instance labels and devise supervised learning approaches. With the ambiguity in instance labels, these studies fail to generalize to the MIL problems with complex structures. To avoid these problems, researchers focus on embedding instance- level information to learn bag representations. In this context, dissimilarity-based representations are known to gen eralize well. This thesis proposes a novel framework in which each bag is represented by its dissimilarities to the prototypes. The framework consists learning mechanisms that provide fast and competitive results compared to the existing distance-based ap proaches on extensive benchmark data sets. The first approach is a simple model that provides a prototype generator from a given MIL data set. We aim to find out prototypes in the feature space to map the collection of instances (i.e. bags) to a dis tance feature space and simultaneously learn a linear classifier for MIL. The second proposal is a tree-based ensemble learning strategy that avoids complex tuning pro cesses and heavy computational costs without sacrificing accuracy. The framework is enriched with the integration of the methods, parameter selection strategy, and en semble design. Furthermore, the proposed methods are extended to the regression domain, namely Multiple Instance Regression (MIR). MIR is a less commonly studied area where the bag labels are real valued data instead of classes. The experiments show that the performances of all proposals are better than the state-of-the art approaches in the literature.
Sat, 01 Jan 2022 00:00:00 GMThttp://digitalarchive.boun.edu.tr/handle/123456789/198002022-01-01T00:00:00ZOptimal server placement, service deployment, and resource allocation in next-generation computer network
http://digitalarchive.boun.edu.tr/handle/123456789/19801
Optimal server placement, service deployment, and resource allocation in next-generation computer network
Ahat, Betül.
With the expansion of mobile devices and new trends in mobile communication technologies, there is an increasing demand for diversified services. To accommodate a large number of services on a common network, it becomes crucial for an operator to optimize resource allocation decisions to satisfy the service requirements in an economical way. In this thesis, the computation architecture design problem is considered first where server placement, service deployment, and task assignment decisions are optimized to maximize the revenue of the operator. The problem is modeled as a mixed-integer linear programming (MILP) formulation and a Lagrangian relaxation- based heuristic algorithm is proposed. Then, the concept of network slicing, which partitions a single physical network into multiple isolated slices, is examined. In the deterministic network slicing problem, the capacities of the computational resources are partitioned into slices each of which is customized for a particular service type. An MILP formulation is presented that takes the delay requirements of services into account. Additionally, two algorithms based on Benders decomposition are devised along with some valid inequalities and cut generation techniques. The problem definition is also extended to consider the stochastic behavior of the service requests. A two-stage stochastic integer programming model is constructed which is then converted into a large-scale MILP model by defining a set of scenarios for the random parameters. A similar decomposition approach is also applied to the stochastic network slicing problem. In our computational study on randomly generated test instances, the validity of our models is assessed and the effectiveness of the proposed solution approaches is demonstrated.
Sat, 01 Jan 2022 00:00:00 GMThttp://digitalarchive.boun.edu.tr/handle/123456789/198012022-01-01T00:00:00ZClassifier combination methods in pattern recognition
http://digitalarchive.boun.edu.tr/handle/123456789/13588
Classifier combination methods in pattern recognition
Baykut, Alper.
This thesis studies methodologies to combine multiple classifiers to improve classification accuracy. Different classifiers, training methods and combination algorithms are covered throughout this study. The classifiers are extended to produce class probability estimates besides their class assignments to be able to combine them more efficiently. They are integrated in a framework to provide a toolbox for classifier combination. The leave-one-out training method is used and the results are combined using proposed weighted combination algorithms. The weights of the classifiers for the weighted classifier combination are determined based on the performance of the classifiers on the training phase. The classifiers and combination algorithms are evaluated using classical and proposed performance measures. It is found that the integration of the proposed reliability measure, improves the performance of classification. A sensitivity analysis shows that the proposed polynomial weight assignment applied with probability based combination is robust to choose classifiers for the classifier set and indicates a typical one to three per cent consistent improvement compared to a single best classifier of the same set.
Tue, 01 Jan 2002 00:00:00 GMThttp://digitalarchive.boun.edu.tr/handle/123456789/135882002-01-01T00:00:00Z