78
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Discipline
Araştırma tabanlı testler baz alınarak model tabanlı testler için modeller geliştirme
Model-based testing facilitates automatic generation of test cases by means of models of the system under test. Correctness and completeness of these models determine the effectiveness of the generated test cases. Critical faults can be missed due to omissions in the models, which are primarily created manually. In practice, these faults are usually detected with exploratory testing performed manually by experienced test engineers. In this thesis, we propose an approach for refining system models based on the experience and domain knowledge of these test engineers. Our toolset analyzes the execution traces that are recorded during exploratory testing activities and identifies the omissions in system models. The identified omissions guide the refinement of models to be able to generate more effective test cases. We applied our approach in the context of two industrial case studies to improve the models for model-based testing of a Digital TV system. After applying our approach, three and four critical faults were detected for the first and second case studies, respectively. These faults were not detected by the initial set of test cases and they were also missed during the exploratory testing activities.
HDTVler için yüksek performanslı düşük karmaşıklıklı gömülü sıkıştırma
HDTV video processors need to keep one or more previously scanned frames as they process streaming video when performing tasks such as frame rate conversion, deinterlacing, and other video enhancement techniques. Reading and writing frames require high bandwidth at HD resolutions. This bandwidth can be reduced by applying compression. Video compression methods do not address this problem as they reduce network traffic while adding extra memory traffic. What is needed is high-performance, low-complexity, and lossless (or near-lossless) image compression. This type of compression method is called Embedded Compression (EC). We propose a novel end-to-end embedded memory compression solution. It can support 4K Ultra HD video streams at 30 Hz with a single core implemented in 180nm ASIC technology, which amounts to a per-core pixel rate twice the competition.
Yazılım tanımlamalı ağlar üzerine çoklu betimlenmiş video çoğa gönderim akışı
Video has become one of the most prominent applications of the Internet. Many of the video streaming applications involve the distribution of content from a CDN source to a large population of interested clients. However, widespread support of IP-Multicast has been unavailable to a large extent due to technical and economical reasons, all stemming from the non-programmable nature of today's Internet. As a solution, streaming multicast video is commonly operated using application level multicast. However, this technique introduces excessive delays for the clients and increased traffic load for the network. This thesis is concerned with the introduction of a SDN based framework that allows the network controller to not only deploy IP-Multicast between a source and subscribers, but also control, via a simple northbound interface, the distributed set of sources where multiple-description coded video content is available. Standard and premium users are envisioned. While standard subscribers are to receive one of the descriptions of the video, premium subscribers will receive multiple descriptions, each from a different source, simultaneously and combine these descriptions prior to playback for increased video quality. In the framework, the controller constructs and maintains a dynamic multicast tree from each source and formulates the associated multicast routes. An experimental testbed has been setup on Mininet to assess the performance of the SDN-based streaming multicast video application using QoS performance metrics on a well-known test videos. We observe that for medium to heavily loaded networks, relative to todays solution of application layer multicast in a non-SDN network, the SDN-based streaming multicast video framework increases the PSNR of the received video significantly, from a level that is practically unwatchable to one that has good quality.
Alana özel uyarlanabilir hizmet gözetimi
We propose an adaptive and domain-specific service monitoring approach to detect partner service errors in a cost-effective manner. Hereby, we not only consider generic errors such as file not found or connection timed out, but also take domain-specific errors into account. The detection of each type of error entails a different monitoring cost in terms of the consumed resources. To reduce costs, we adapt the monitoring frequency for each service and for each type of error based on the measured error rates and a cost model. We introduce an industrial case study from the broadcasting and content-delivery domain for improving the user-perceived reliability of Smart TV systems. We demonstrate the effectiveness of our approach with real data collected to be relevant for a commercial TV portal application. We present empirical results regarding the trade-off between monitoring overhead and error detection accuracy. Our results show that each service is usually subject to various types of errors with different error rates and exploiting this variation can reduce monitoring costs by up to 30\% with negligible compromise on the quality of monitoring.
Açık dünya politika muhakemesi için akıllı bilgi toplama
Policies play an important role in autonomous multi-agents systems where each agent aims to achieve its own goals. Policies and related mechanisms allow authority or society to regulate the actions of agents to prohibit malicious and undesirable activities. Without policies, society could be harmed by irresponsible and malicious activities of its members. On the other hand, reasoning with policies is not trivial; it requires extensive knowledge about the environment. If the knowledge is incomplete or missing, reasoning with policies may not be possible. In this thesis, we propose a proactive approach for gathering information to reason with policies. While our approach can be used in various settings, we provide two case-studies; one in social networking domain and the other in on-line advertisement domain. Through experiments we demonstrated that our approach allows high rate of success during policy reasoning when the knowledge bases is not complete.
Eylem anlama ve duyguların beyin mekanizmalarına hesapsal yaklaşımlar
Through evolution living beings have gained unique features to deal with apparently easy but computationally expensive problems such as mate selection, learning sensorimotor skills and decision making. Thus, understanding how a biological system can process sensory information, interpret the probable results and find a solution in relatively short time to faced problems have become an attractive research area for computational neuroscience, artificial intelligence (AI) and robotics. In this thesis we focused on mirror neurons in the ventral premotor cortex (area F5) and the functional aspects of emotions from a computational but biologically plausible way. In the former part, the raw neural firing data from area F5 of macaque monkeys are analyzed to undercover neural representation using a decoding framework. For this, we propose two methods to detect mirror neurons by using machine learning and statistical analysis techniques. In the later part, we present that higher level emotions (those that have putatively evolved after the basic emotions of fear, anger etc.) are the behavioral manifestation of self-regulation mechanisms of computational (neuronal) energy expenditure for cognitive processing. To realize this proposal, we chose a tractable computational mechanism that may be considered as a model of neural computation mechanisms of the brain and deploy it on a robotic platform (Darwin-OP).
Dinamik hareket birimleri ile insan hareketi tanıma
Dynamic Movement Primitives (DMPs)-originally a method for movement trajectory generation has been also used for recognition tasks. However there has not been a systematic comparison between other recognition methods and DMPs using human movement data. We have implemented a movement recognition method based on DMPs with Gaussians centered equally spaced in phase variable and scaled one-nearest-neighbor weight comparison. Furthermore, in thesis, we presented a comparison of commonly used Hidden Markov Model (HMM) based recognition with our implementation of DMP based recognition using human generated letter trajectories. As the working principles of these two methods are very different, in addition to the performance, the numbers of adaptable parameters that are used in each method and, process time were compared. The results indicate that DMP gives better results than HMM in the tests with noiseless data, noisy data and derogated data with given human movement dataset.
Etkin beceri sentezi için eşzamanlı insan-robot öğrenmesi
It is generally expected that robots and autonomous agents will become a part of our daily lives in the coming decades. However, it is not feasible to program robots in advance for all possible tasks using classical robot programming. Therefore, intuitive and easy robot programming is one of the active research areas in robotics. We propose and implement a human-in-the loop robot skill synthesis that involves simultaneous adaptation of the human and the robot. In this framework, the human demonstrator learns to control the robot in real-time to make it perform a given task. At the same time, the robot learns from the human guided control creating a non-trivial coupled dynamical system. The research question we address is how this system can be tuned to facilitate faster skill transfer or improve the performance level of the transferred skill. At the beginning of the skill transfer session, the human demonstrator controls the robot exclusively as in teleoperation. As the task performance improves the robot takes increasingly more share in control, eventually reaching to full autonomy. The proposed framework is implemented and shown to work on some tasks such as physical cart-pole setup, cart-pole balance simulation, and mountain car. To assess whether simultaneous learning has advantage over the standard sequential learning (where the robot learns from the human observation but does not interfere with the control) experiments with two groups of subjects were performed. Moreover, reinforcement learning is applied to model a human demonstrator to verify simultaneous framework. The results indicate that the final autonomous controller obtained via simultaneous learning has a higher performance in the mentioned tasks.
PL/SQL programları için otomatik prosedür gruplama
Large software systems have to be decomposed into separate, modular units for providing appropriate abstractions and improving maintainability. There exist clustering techniques that are applied to provide such abstractions by automatically grouping system modules based on dependencies among them. Hereby, dependency is usually measured as the extent to which a module refers to elements of another module. This approach cannot be directly applied for all types of programs. Some programs involve modules that are indirectly coupled. For instance, PL/SQL programs include procedures that are in most cases coupled due to their database operations although they do not make calls to each other. In this thesis, we provide an approach and a tool that supports automated modularization of software systems by considering this type of dependencies We also extend this approach for multiple, different types of dependencies. We construct several dependency matrices each of which captures a different type of dependency among the system modules. First, we perform clustering according to each of these matrices separately. Then, we perform cluster aggregation (meta clustering) on the obtained clustering results to propose a packaging structure to the designer. We performed two industrial case studies on real PL/SQL programs from the telecommunications domain. Many unexisted packages were proposed by our tool and the accuracy of the results were confirmed by domain experts.
Boolean fonksiyonların işaret gösterimindeki terimlerin sıfırlanma düzenleri
Boolean functions (BF) are one of the fundamental concepts in discrete mathematics. It is possible to represent any BF by a unique polynomial when one takes -1 as True and 1 as False. Coefficients of the polynomial representing the given BF can be found with Lagrange interpolation. When the exact interpolation criterion is replaced with the signmatch criterion, one can find infinitely many sign representing polynomials for a given truth table. The problem of finding a minimum number of monomial set that is sufficient to represent a BF is a difficult mathematical problem. This thesis aims to contribute to its solution by investigating the zeroability patterns of monomials. To this end, we asked which monomials must be in a minimum sign representing polynomial. This question drove us to make numerical investigations on the BFs in lower dimensions. For all 3- and 4-variable BFs, we found all the monomial subsets, whose elements can be zeroed and we introduced a graph representation indicating whether particular pairs of monomials could be absent from any sign representation. In addition to the numerical investigations, we have also proved that if a three-element monomial set S, could not be absent altogether from the sign representation of a BF, then there must be at least a two element subset of S which could not be absent in any sign representation of that BF. We expect these results will give support to the development of heuristic algorithms to construct close-to-minimum number of monomial sign representing polynomials for BFs.
Sinerjik görev uygulaması için insan-robot işbirliği
There is great potential for human and robot to work together as a team, since this collaboration can take advantage of both human and robot capabilities, cover their weakness and yield a higher performance. We propose and implement a human-robot collaboration framework where, while the human tries to perform a task, the robot infers the human intention and assists the human in achieving the inferred goal. We explore how the human is influenced when (s)he interact with machine autonomy, and whether there is any advantage in task performance when human shares control with an autonomous agent. In particular, we investigate whether interacting with autonomy can aid humans to improve their performance in shorter time. We realized this collaboration system by designing a ball balancing task in which the goal is to move and balance the ball on a target position on a tray held by a robotic arm. The human performs the task by controlling the robotic arm with an interface which tilts the tray and moves the ball while the robot infers the target ball position by observing the trajectory of the ball, and augments the human control commands for assisting in task execution. The length of ball movement trajectory, completion time and positional error were chosen as the measures to evaluate the task performance. To assess the impact of our system on human learning and task execution a set of experiments were conducted under two conditions, human control condition where human performs the task alone and share control condition where both human and robot are involved in performing the task. 20 naive subjects were volunteered to perform the experiment in four continuous days. The result of these experiments suggests that not only the task execution can be improved in collaboration with robot compare to when the humans perform the task alone but also this collaboration system can make the human learning to progress faster.
Robotik bir el üzerinde çevik manipülasyon
In robotics, flexible and the dexterous manipulation are one of the most desired type of skills. To this end, we investigate dexterous manipulation skills on an anthropomorphic robot hand. In the first part of the study, a sensorless grasping method is described. Although the high-precision sensing is highly relevant for precise grasps, precision is often not necessary to perform power grasps. An alternative approach is proposed for robotic grasping tasks based on external force estimation. Estimation accuracy is confirmed using a force sensor and the estimations are found to be useful for creating soft/power grasp behavior. In the second part, human-in-the-loop heterogeneous control for dexterous manipulation is investigated on a setup with a robotic hand and a robotic arm. The goal of the study is to experimentally verify that in tasks where the manual and explicit trajectory tuning is not possible, the autonomous movement can be learned by giving a basic policy to a robotic system, after which a human can learn and transfer an orthogonal complex part of the policy. The approach is shown on a ball swapping task in which a robotic arm is controlled by the human and a robotic hand is given an initial basic policy. In the results, we experimentally show that, in certain tasks, complex autonomous policies can be constructed by delegating the complex learning part to a human, the simple part to an autonomous agent, finally creating an autonomous control policy by recombining the parts.
Çok ölçekli ikili benzerlik yüz tanıma için yerel ikili örüntü varyantı
Face recognition problem has been studying for more than four-decade, and many descriptors and neural network architectures were proposed. The aim is simple, extract features from the same subjects for training and test face image sets, if the proposed method was accurate, the extracted features categorized under the same label. However, the problem starts with the illumination effect on the images; the illumination effect may cause the extracted features for the same subject to be classified with the different labels. Therefore, illumination and other environmental impacts should be removed for accurate classification. One solution for eliminating environmental effect is using Local Binary Pattern (LBP) descriptor. LBP is an illumination invariant, computationally simple, and highly discriminative visual descriptor. Therefore, LBP based descriptors have been developing for more than a two-decade for solving face recognition problem. LBP′s computationally simple property make it applicable to different types of computer vision problems, also there are many examples of LBP variants either achieved state-of-the-art results in a particular application or complementary to the LBP. Having been inspired from the results, in this thesis, an LBP variant descriptor, Multi-scale Binary Similarity approach is proposed. MSBS encodes face image characteristic by analyzing the pixel relationships in selected components. The encoded features of the MSBS trained with Support Vector Machines (SVM) and tested with AT&T, Extended Yale B, Georgia Tech and MNIST datasets. The results show that MSBS outperforms most of the proposed approaches in the literature.
Nesnelerin interneti için semantik düzenleyici kurallar sistemi
With the proliferation of technology, connected and interconnected devices (hence- forth referred to as IoT) are fast becoming a viable option to automate the day-to-day interactions of users with their environments. However, with the explosion of IoT de- ployments we have observed in recent years, manually managing the interactions between humans-to-devices, and especially devices-to-devices, is an impractical task, if not an impossible task. This is because devices have their own obligations and prohibitions in context, and humans are not equipped to maintain a bird's-eye-view of the interaction space. Motivated by this observation, in this thesis, we propose a semantic policy framework that (a) supports representation of high-level and expressive user policies to govern the devices and services in the environment; (b) provides efficient procedures to refine and reason about policies to automate the management of interactions; and (c) delegates similar capable devices to fulfill the interactions, when conflicts occur. We then describe how to combine ontology-based policy reasoning mechanisms with in-use IoT applications to customize and automate device behaviors and discuss how the policy framework can be extended with data federation to handle diverse and distributed data sources. We demonstrate that smart devices and sensors can be orchestrated through policies in diverse settings, from smart home environments to hazardous workplaces, such as coal mines. Lastly, we evaluate our approach using real applications with real data and demonstrate that our approach is scalable under high load of data and devices.
A unified framework for benchmarking sparse matrix-vector multiplication methods
Sparse matrix-vector multiplication (SpMV) is an important sparse linear algebra kernel that has a wide range of application domains, including computational science, graph analytics, machine learning and many more. Due to its significance, numerous studies have been conducted and are still being proposed to improve the performance of SpMV. Most of the studies evaluate the performance of their method in a custom experimental environment, which weakens the reproducibility of the empirical results, and also makes it hard to compare the proposed method to a wide range of existing methods. In this study, we address this problem by introducing an easy-to-integrate benchmarking framework that is able to unify SpMV methods in a single experimental environment to obtain consistent evaluation results. As a proof-of-concept, we have integrated several state-of-the-art CPU and GPU-based SpMV methods in our framework. We make the framework available as an open-source software for the convenience of researchers.
Finite element analysis in a cloud computing environment
In this thesis, the challenges faced and lessons learned while establishing a large-scale high performance cloud computing service that enables online mechanical structural analysis and many other scientific applications using the finite element analysis (FEA) technique, will be described. Within an High Performance Computing (HPC) environment, several jobs with different demands can co-exist thus it becomes a challenge for the service provider to efficiently utilize its own resources while also satisfying the quality expectations of job submitters. Such a service is intended to process many independent and loosely-dependent tasks concurrently. In order to reach optimal job scheduling metrics each job type that can be submitted to the cluster must be carefully examined, its space and time characteristics must be well-understood and quantified. Challenges faced include accurate characterization of complex FEA jobs, handling of many-task mixed jobs, sensitivity of task execution to multi-threading parameters, effective multi-core scheduling within a single computing node, and achieving seamless scaling across multiple nodes. It is found that significant performance gains in terms of both job completion latency and throughput are possible via dynamic or "smart" batch partitioning and resource-aware scheduling compared to the naive Shortest Job First (SCF) and aggressively-parallel scheduling techniques. Chapter 3 of this thesis present an end-to-end discussion on the technical issues related to the design and implementation of a new cloud computing service for finite element analysis (FEA). Several design choices for HPC services at different layers of the cloud computing architecture are investigated to simplify and broaden its use cases. Investigations start with the software-as-a-service (SaaS) layer and compare parallel linear equation solvers. In order to minimize job latency and maximize the overall job throughput, several matrix characteristics are perceived. Developing such an understanding is also crucial for HPCaaS systems to automatically select the amount of computing resources per job. In following sections, the design of a ''smart'' scheduler that can dynamically select some of the required parameters, partition the workload and schedule it in a resource-aware manner will be demonstrated. Results showing that an up to 7.53x performance improvement over an aggressive scheduler using mixed FEA loads, will be presented. In addition to the performance studies, a complementary discussion on critical issues related to the data privacy, security, accounting, and portability of the cloud service will also be given. The new trend in engineering is to solve complex computational problems in the cloud over HPC services provided by different vendors. To further deepen the analyses of workloads representing HPC-related tasks in science and engineering, in chapter 4, performances of direct vs. iterative linear equation solvers are compared to help with the development of job schedulers that can automatically choose the best solver type and tune them (e.g. precondition the matrices) according to job characteristics and workload conditions that are frequently encountered on HPC cloud services. As a proof of concept, three classical elasticity problems will be used, namely a Cantilever beam, Lame problem and Stress Concentration Factor (SCF). These models theoretically represent many real-life mechanical situations in structural engineering, namely aerospace, automotive, construction and machinery industries. The representative linear problems are meshed with increasing granularities, which leads to various matrix sizes; largest having 1 billion non-zero elements. Detailed finite element analyses over an IBM HPC cluster are executed. First, a multi-frontal parallel is used, sparse direct solver and evaluate its performance with Cholesky and LU decompositions of the generated matrices with respect to memory usage, and multi-core, multi-node execution performances. As for the iterative solver, the PETSc library is used and carried out computations with several Krylov subspace methods (CG, BiCG, GMRES) and preconditioner combinations (BJacobi, SOR, ASM, None). Later in Chapter 4, the direct and iterative solver results are compared and contrasted in order to find the most suitable algorithm for varying cases obtained from numerical modeling of these three-dimensional linear elasticity problems. In addition to aforementioned studies, as a supplementary research, infrastructure-as-a-service (IaaS) layer for HPC is examined and characteristics like application performance, load isolation, and deployment speed issues using application containers (Docker) are observed. These characteristics are also compared to physical and virtual machines (VM) over a public cloud. For this purpose, HPC-specific deployment using application containers technology is evaluated and performance metrics are examined in order to contribute to evaluation of these technologies for job schedulers to be used on Cloud Computing infrastructures. This phase of the research focuses on the understanding the behavior of cloud computing infrastructures under circumstances where deployment and utilization of containers (Docker) with a chosen software is necessary. To summarize, this multi-disciplinary doctoral thesis covers most of the critical aspects and computational challenges of providing FEA in the cloud for structural mechanics including ease of deployment, batch-level performance, job-level isolation, financial accounting and content security. It utilizes several modern software tools and techniques, while also contributing new ones to the literature.
Scalable analysis of large-scale system logs for anomaly detection
System logs provide information regarding the status of system components and various events that occur at runtime. This information can support fault detection, diagnosis and prediction activities. However, it is a challenging task to analyze and interpret a huge volume of log data, which do not always conform to a standardized structure. As the scale increases, distributed systems can generate logs as a collection of huge volume of messages from several components. Thus, it becomes infeasible to monitor and detect anomalies efficiently and effectively by applying manual or traditional analysis techniques. There have been several studies that aim at detecting system anomalies automatically by applying machine learning techniques on system logs. However, they offer limited efficiency and scalability. We identified three shortcomings that cause these limitations: i)Existing log parsing techniques do not parse unstructured log messages in a parallel and distributed manner. ii)Log data is processed mainly in offline mode rather than online. That is, the entire log data is collected beforehand, instead of analyzing it piece-by-piece as soon as more data becomes available. iii)Existing studies employ centralized implementations of machine learning algorithms. In this dissertation, we address these shortcomings to facilitate end-to-end scalable analysis of large-scale system logs for anomaly detection. We introduce a framework for distributed analysis of unstructured log messages. We evaluated our framework with two sets of log messages obtained from real systems. Results showed that our framework achieves more than 30% performance improvement on average, compared to baseline approaches that do not employ fully distributed processing. In addition, it maintains the same accuracy level as those obtained with benchmark studies although it does not require the availability of the source code, unlike those studies. Our framework also enables online processing, where log data is processed progressively in successive time windows. The benefit of this approach is that some anomalies can be detected earlier. The risk is that the accuracy might be hampered. Experimental results showed that this risk occurs rarely, only when a window boundary cross-cuts a session of events. On the other hand, average anomaly detection time is reduced significantly. Finally, we introduce a case study that evaluates distributed implementations of PCA and K-means algorithms. We compared the accuracy and performance of these algorithms both with respect to each other and with respect to their centralized implementations. Results showed that the distributed versions can achieve the same accuracy and provide a performance improvement by orders of magnitude when compared to their centralized versions. The performance of PCA turns out to be better than K-means, although we observed that the difference between the two tends to decrease as the degree of parallelism increases.
Clothing image retrieval with triplet capsule networks
Clothing image retrieval has become more important after some major developments in Computer Science and the emergence of e-commerce. Recent studies generally attack this problem by using Convolutional Neural Networks (CNNs). Despite their popularity, CNNs, by nature, have some intrinsic limitations such as losing the hierarchical spatial relationship between the parts of an image, and not being robust to affine transformations. Most recently proposed network architecture, namely Capsule Networks, has the ability to overcome these limitations by preserving the part-whole relationship and pose information in the images. In this thesis, we investigate in-shop clothing retrieval performance of densely-connected Capsule Networks with dynamic routing. To achieve this, we propose Triplet-based designs of Capsule Network architecture with two different feature extraction methods: Stacked-convolutional (SCCapsNet) and Residual-connected (RCCapsNet) Capsule Networks. Experimental results of our proposed designs on in-shop clothing retrieval show that SCCapsNet achieves 32.1\% Top-1, 81.8\% Top-20, and 90.0\% Top-50 recall-at-K scores; whereas RCCapsNet has even better performance with 33.9\% Top-1, 84.6\% Top-20, and 92.6\% Top-50 recall-at-K scores. These figures demonstrate that both of our designs outperform the baseline study and the earlier approaches by a wide margin without using any extra supportive information besides to the images. Moreover, when compared to the SOTA architectures on clothing retrieval, our proposed Triplet Capsule Networks achieve comparable recall rates with only half of the parameters used in the SOTA architectures. In the future, our designs may inherit extra performance boost due to advances in the relatively new Capsule Network research.
Teşvik odaklı ve mahremiyete dayalı bilgi paylaşımı için etmen temelli müzakere
While customizing their services, companies usually use their users' data. According to the new regularization, it is required to get the permission of their users to be able to store and share their users' private data. The current approaches rely on requesting access rights by providing some incentives. The customers can only accept or reject the possible incentive offered by the companies exchange for giving access rights. This thesis introduces an agent-based, incentive-driven, and privacy-preserving information sharing framework. One of the main contributions of this thesis is to give the data provider agent an active role in the information sharing process and to change the currently asymmetric position between the provider and the requester of data and information (DI) to the favor of the DI provider. Instead of a binary yes/no answer to the requester's data request and the incentive offer, the provider may negotiate about excluding from the requested DI bundle certain pieces of DI with high privacy value, and/or ask for a different type of incentive. We show the presented approach on a use case and conduct a user experiment. Questionnaire responses showed that participants like the idea of negotiation on their information sharing policies with the companies. Furthermore, this thesis proposes an acceptance strategy using deep reinforcement learning for automated negotiating agents. In the automated negotiation literature, most of the acceptance strategies are based on some predefined rules. In contrast, this thesis proposes to use reinforcement learning in order to learn when to accept opponent's offer. Our experimental evaluation shows that the developed acceptance strategy performed as well as AC-Next acceptance strategy.
Embedding-based clustering for target specific stances
We propose an unsupervised user stance detection method to capture fine grained divergences in a community across various topics. We employ pre-trained universal sentence encoders to represent users based on the content of their tweets on a particular topic. User vectors are projected onto a lower dimensional space using UMAP, then clustered using HDBSCAN. Our method performs better than previous approaches on two datasets in different domains, achieving precision and recall scores ranging between 0.89 and 0.97. We compiled a dataset of more than 300k tweets about UEFA Super Cup's 2019 final, and tagged 12k users as Liverpool FC or Chelsea FC fans. We utilized our method to analyze the stances of Twitter users noting a correlation between user stances towards various polarizing issues. We used the resultant clusters to quantify the polarization in various topics, and analyze the semantic divergence between clusters.
A novel method for automatic fish counting based on machine learning using morphological features
This thesis offers a machine learning based computer vision approach for counting fish in a non-stop running narrow water stream (or equally on a conveyor belt or similar mechanism) in a fish farm. We call this "continuous operation". Such setup requires that counting is done extremely fast, which we call "real-time" operation. To our knowledge, our fish counting solution is the only such method, namely, continuous and real-time. The hardest subproblem here is correctly counting "overlapped fish" and yet be flexible in terms of fish sizes. Most of the previous works are not real-time due to their inefficient approach to counting overlapped fish. Our superiority in speed mostly stems from selecting fewer and computationally less expensive morphological features (which we feed to machine learning). Although the state-of-the-art is pretty accurate (98.9%), we were able to improve it to 99.4% for even very difficult test cases. Since our solution is continuous unlike previous work in the literature, we had to also solve the problem of "merging fish" (overlapped or not) lying at the boundary of consecutive frames.
Familiarity effect on human-agent negotiations
Artificial Intelligence has changed our world in various ways. People have started to interact with a variety of intelligent systems on a daily basis. As the interaction between human and AI systems increases day by day, the factors influencing their communication have become more and more important especially in the field of human-agent negotiation. Therefore, it is necessary to study the factors affecting human-human negotiation while designing agents negotiating with their human counterparts. As familiarity is one of these factors, this work aims to investigate the effect of familiarity on human-agent negotiation so that we can design more effective negotiation systems. Being familiar to other party may influence how we interact and hence the process and outcome of the negotiation. Our hypothesis is that negotiating with a familiar opponent would create a difference in terms of negotiation process and outcome. In order to study this effect in human-agent negotiations, we developed negotiation framework in which human participants negotiate with an avatar in a bilateral fashion. To measure the effect of the appearance familiarity in negotiation, two control variables are defined: negotiating with a celebrity avatar and negotiating with a non-celebrity avatar. In order to avoid the learning effect, we adopt a between-subject experiment design. We recruited 67 participants and analyzed their negotiation data elaborately as well as their subjective opinions specified in the questionnaires. Our experimental results showed that being familiar with the opponent affected both negotiation process and outcome. Particularly, human participants had a tendency to be more collaborative when their opponent is a celebrity avatar versus a non-celebrity avatar.
X-ray analysis prediction of BGA components in PCBA production with neural networks
Printed circuit boards are the most important part of all electronic devices used today, and the production of these boards consists of many critical processes. Assembly production lines include different inspection machines such as Solder Paste Inspection (SPI), Automatic Optical Inspection (AOI), X-Ray Inspection Device to detect defects and misplaced components on the circuit boards. SPI and AOI machines determine these problems by checking that different measurement values remain between certain threshold values, but these machines cannot provide a perfect detection mechanism since threshold values are determined by a maintenance technician by trial and error. Therefore, the circuit boards are also controlled by X-Ray Inspection machines, thus the solder paste that cannot be inspected visually by human can be easily inspected. This project aims to create a neural networks model that uses SPI and AOI measurement parameters as input features and predicts potentially defective boards before the boards go through X-Ray inspection. This model allows suspicious circuit boards to be tested in X-Ray Inspection machine, instead of testing of some circuit boards randomly. Thus, the number of X-ray devices required to test the boards is reduced, the circuit boards to be selected for sampling are selected from the boards that are likely to be defective rather than randomly selected. By virtue of this model, suspicious solders and components will be marked on the boards that will be sent to X-Ray test station as suspicious and the operator will be prevented from missing the defects that may occur in these areas. It is foreseen that this system will also identify contact without connection defects, also known as head-in-pillow defects, which cannot be detected in SPI, AOI and Functional Verification Test (FVT) processes, thus the quality of the boards produced is likely to increase. The data will be collected for 17MB170R4 model circuit boards, which are the main board of televisions, produced in one of the production lines at Vestel Electronics Factory, and the model will be applied for four BGA elements on this board. In this project, it is aimed to work on BGA elements because the solder areas of these elements cannot be inspected visually after the component is placed on the board. Due to the model, it is possible to detect most of the defective boards by applying X-Ray testing to only 1% of all circuit boards produced.
Uncertainty assessment for speaker verification systems using a bayesian approach
The Automatic Speaker Verification (ASV) systems are developed to discriminate the genuine speakers from the spoofing attacks and they are also used as a security application in various industries (e.g., Banking and telephone-based systems). The spoofing countermeasure systems (SCS) are important for the ASV systems to protect themselves against spoofing attacks. In general, the SCSs are developed using the cross entropy loss function and the softmax classification layer to perform the best classification scores. Even though the softmax function is popularly used as a classification layer for the deep neural network tasks, it increases the uncertainty of the estimated class probabilities by squishing the probabilistic predictions of the predictive models. The aim of this work was to decrease uncertainty of the conventional cross entropy metrics and softmax function SCS by using the Bayesian approach. To accomplish this, multiple SCSs were developed to outperform the base system of the Automatic Speaker Verification Spoofing and Countermeasures 2017 Challenge. The Bayesian approach was applied to the best model (e.g., the model which performed the lowest EER score) to decrease the uncertainty of the conventional cross entropy metrics and softmax function SCS. The uncertainty of the both systems were compared with the probability distribution function, AUC value and the ROC curve. As it can be observed from the ROC curve, the Bayesian network decreased the uncertainty of the conventional cross entropy metrics and softmax function SCS by increasing AUC value 14%. Also the Bayesian network has provided the lowest EER score (16.79%) by outperforming the base system of the ASV spoof 2017 challenge.
Sparse voxel based 3D object detectionfrom RGB-D data
Accurate and fast 3D object detection plays a role of paramount importance for safe and capable autonomous machines. LiDAR point cloud based methods have demonstrated impressive results, yet expensive LiDAR sensors make such approaches infeasible for wide-scale adaptation. Camera based methods on the other hand are performing sub-optimally given safety and accuracy requirements. Traditionally, camera based 3D object detection is performed by generating pseudo-LiDAR point clouds from RGB-D data and using point-cloud based methods, however, irregular nature of point cloud data representation makes it challenging to exploit spatial local correlations on 3D space and point cloud based methods generally suffer from this. We propose Sparse Voxel based 3D Object Detection, our approach differs from traditional approaches by converting point cloud information to sparse voxel grid and utilizing sub-manifold sparse convolutions to extract information instead of PointNet based models. Our approach not only outperforms its point-cloud based counterparts with a wide margin but also comes with the advantage of being efficient to compute.
Otomatikleştirilmiş ikili pazarlıklarda ilişkisel ve frekansçı rakip modelleme yaklaşımları
This thesis mainly focuses on the problem of learning opponent's preferences during the negotiation in bilateral automated negotiation in which agents negotiate with each other to reach an agreement. Accordingly, it addresses the problems with the classical frequentist approach and advances the state-of-the-art in opponent modeling in automated negotiation by introducing a novel frequency opponent modeling mechanism, which updates some of the assumptions introduced by classical frequency approaches. Moreover, this thesis also proposes adopting association rule mining techniques to learn the opponent's preferences in bilateral negotiation. An extensive evaluation of those proposed approaches shows that the proposed approaches outperform the classical frequency model. In addition, this thesis argues that while optimizing one's utility function is essential, agents in a society should not ignore the opponent's utility in the final agreement to improve the agent's long-term interests in the system. It aims to show whether or not it is possible to design a social agent (i.e., one that aims to optimize both sides' utility functions) while performing efficiently in an agent society. Accordingly, we propose a social agent supported by a portfolio of strategies, a novel tit-for-tat concession mechanism, and a frequency-based opponent modeling mechanism capable of adapting its behavior according to the opponent's behavior and the state of the negotiation. The results show that the proposed social agent does not only maximize social metrics such as the distance to the Nash bargaining point or the Kalai point but also is shown to be a pure and mixed equilibrium strategy in some realistic agent societies.
Acil sağlık sistemleri için blockchain tabanlı güvenlik mekanizması
Elektronik sağlık kayıtları (ESK'ler) günümüz sağlık endüstrisinde çok önemli bir rol oynamaktadır, bu kayıtlar güvenlik ve gizliliğin ihlaline açık hassas ve özel sağlık veri varlıklarını içermektedir. Bu olası veri ihlallerinin hastanın mahremiyetinin ihlali, ESK'ye yetkisiz erişim, veri değişikliği ve hastanın hayatını tehlikeye atma gibi birçok sonucu olabilir. Son zamanlarda önerilen ESK sistemleri, hastanın güvenliğini korumak için güçlü güvenlik özellikleri ile birlikte geliyor. Ancak, Kişisel Sağlık Kayıtlarına (KSK'ler) erişim kontrolü yönetimi ile ilgili hala sorunlar bulunmaktadır. Daha önce sunulan sistemlerde, sisteme erişimi kontrol etmede hasta ana rolü oynar, ve bu acil durumlarda bir belirsizliğe yol açarken, hasta herhangi bir erişim izni veremez. Bu çalışmada, Hyperledger Sawtooth blockzincir ile tasarlanan güvenli ve özel bir erişim kontrol mimarisi sunarak acil durumlarda tıbbi sağlık kayıtlarının (TSK'ler) tutulması için yeni bir çerçeve öneriyoruz. Blok zincirinin benzersiz özelliklerinden yararlanan sistemimiz, acil bir senaryoda hastanın tıbbi verilerine kısa sürede güvenli bir erişim sağlar. Simülasyonumuzun sayısal sonuçları, yanıt süresi, bellek tüketimi, verim, genel gizlilik ve güvenlik açısından benzer sağlık sistemleriyle karşılaştırıldığında önerilen mimarimizin performansının daha iyi olduğunu ve kullanılabilirleğini gösterir.
Motor kontrol ve beynin bilişsel karar verme mekanizmalarını analiz etmek üzere tersine pekiştirmeli öğrenme ile keşifler
Reinforcement Learning is a framework for generating optimal policies given a task and a reward/punishment structure. Likewise, Inverse Reinforcement Learning, as the name suggests, is used for recovering the reasoning behind an optimal policy based on demonstrations from an expert. We set out to explore whether recent Reinforcement Learning and Inverse Reinforcement Learning methods can serve as a computational tool for investigating optimality principles of motor control and cognitive decision-making mechanisms of the brain. For this purpose, we have targeted several different tasks involved with different parts of the sensorimotor learning mechanism of the brain. We aim to recover the optimality principles employed by the brain for various control and decision-making tasks. If this is achieved, we can analyze, understand, mimic and improve demonstrated behavior with less bias, which we hope is a step forward in understanding the process of learning in both human-based and artificial systems. For the scope of this thesis, we have evaluated two tasks. The first task was investigating the applicability of perceptual development for Reinforcement Learning. For this task, we have proposed a perceptual development based learning regime for a Reinforcement Learning agent, and the results obtained suggest that a suitable perceptual development regime may improve the learning progress and yield better-performing agents. The second task was to predict reward function parameters of a provided trajectory in a standing up under perturbation scenario. For this task, we have proposed two different Inverse Reinforcement Learning approaches. Our results indicate that we were able to infer valid reward parameters on synthetic data.
Kendinden denetimli derin öğrenme ile multispektral görüntü eşleştirme
This thesis presents a self-supervised deep learning framework for multispectral image matching, addressing the core challenges posed by nonlinear radiation distortions (NRDs), viewpoint variations across spectral modalities, and the scarcity of annotated datasets. Existing methods often exhibit strong modality dependence and rely heavily on costly supervision, such as depth maps or calibrated camera poses, thereby limiting their generalizability across diverse spectral domains. The framework introduces an improved self-supervision strategy—Improved Multispectral Homographic Adaptation—that enhances pseudo ground truth keypoint generation in cross-spectral settings while ensuring invariance to viewpoint changes. By incorporating a spectrum-aware windowing rule, this method increases both the repeatability and the density of detected keypoints under spectral differences, thereby improving matching performance. This enhancement ultimately leads to more accurate multispectral image registration and is validated on UAV-acquired visible–thermal datasets. Building on this self-supervision strategy, the XPoint framework is proposed as a modular and fully self-supervised image matching architecture. It integrates a pretrained VMamba encoder for robust, modality-invariant feature extraction, alongside lightweight decoder heads for keypoint detection, feature description, and homography regression. This design enables efficient, label-free learning from aligned image pairs and facilitates rapid adaptation across diverse spectral modalities. The framework is designed to be scalable and easily adaptable, requiring no additional supervision beyond image pair alignment. The approach is evaluated across five public benchmarks spanning VIS-TH, VIS-NIR, VIS-LWIR, VIS-SAR datasets. Experimental results demonstrate competitive or superior performance in feature matching and multispectral image registration tasks, while maintaining high computational efficiency. This progression—from spectrum-aware self-supervision to a generalizable matching framework—positions XPoint as a practical solution for real-world multispectral applications, particularly in environments characterized by limited supervision and high spectral variability.
Değişen kaynak dosyalarını daha önce çözülen hatalar ile ilişkilendirerek test senaryosu seçimi
Software continuously changes to address new requirements and to fix defects. Regression testing is performed to ensure that the applied changes do not adversely affect existing functionality. The increasing number of test cases makes it infeasible to execute the whole regression test suite. Test case selection is adopted to select a subset of the test suite, which is associated with the changed parts of the software. These parts are assumed to be error-prone. We present and evaluate a test case selection approach in the context of black-box regression testing of embedded systems. In this context, it is challenging to relate test cases with a set of distinct source code elements to be able to select those test cases associated with the modified parts of the source code. We analyze previously fixed defects for this purpose. We relate test cases that detect these defects with the source files that are previously modified for fixing them. Then, we select test cases related with source code files that are modified in the subsequent revision. The strength of this relation is determined as the number of changes associated with fixed defects previously detected by the same test cases. We conduct a case study on 3 real projects from the consumer electronics domain. Results show that it is possible to detect from 65% up to 85% of the defects detected by the whole test suite by selecting from 30% up to 70% of the test cases.
Kara listeleme ve güven puanlama kullanan ıot ağları için bir güvenlik protokolü
There have been a number of high-profile incidents to compromise and attack larger networks of IoT devices, drawing attention to the need for IoT security. The purpose of IoT security is to ensure the availability, confidentiality, and integrity of IoT networks. However, due to the heterogeneity of IoT devices and the possibility of attacks from both inside and outside the network, securing an IoT network is a difficult task. Handshake protocols are useful for achieving mutual authentication which allows secure inclusion of devices into the network. However, they cannot prevent malicious network-based attacks once attackers enter the network. Use of autonomous anomaly detection and blacklisting prevent nodes with anomalous behavior from joining, re-joining, or remaining in the network. This is useful for securing an IoT network from insider network-based attacks. Similarly, trust scoring is another popular method that can be used to increase the resilience of the network against behavioral attacks. The contributions of this thesis are threefold. First, we propose a new handshake protocol that can be used in device discovery and mutual authentication to ensure the security of the IoT network from outsider attacks. In the proposed handshake protocol, a Physical Unclonable Function (PUF) is utilized for the session key generation to reduce computational complexity. The proposed protocol is resilient to Man-in-the-middle, replay and reforge attacks as proven in our security analysis. Secondly, we propose a machine learning (ML) based intrusion and anomaly detection to prevent network-based attacks from the insiders. Finally, we propose a trust system which utilizes blockchain for managing the trust of a dynamic IoT network to increase resilience against behavioral attacks. Simulation results show that the proposed comprehensive security framework is capable of ensuring the security of an IoT network from both inside and outside attackers.
Yayın akışını geliştirmek için sunucu ve istemci tarafı algoritmaları
HTTP adaptive video streaming is a technique widely used on the internet today to stream live and on-demand content. Server and client-side algorithms play an important role in improving user experience in terms of metrics such as latency, rebufferings and rendering quality. After explaining the commonly used metrics, we analyzed four main aspects of video streaming (i) bandwidth prediction accuracy, (ii) utilization of playback speed, (iii) adaptive streaming for content-aware-encoded videos, and (iv) head motion awareness for 360-degree videos. 360-degree video streaming requires much higher bandwidth compared to conventional video streaming. We demonstrate that most of the algorithmic improvements achieved for video streaming can also be applied to Viewport Dependent Streaming (VDS) for 360-degree videos. It is also important that in 360-degree video streaming, we have a Head Mounted Display (HMD) device that is capable of pointing the viewport orientation of the user. We also investigate and improve the rate-adaptation algorithms for 360-degree videos by developing several new algorithms making use of the HMD. The new algorithms proposed in this thesis are Low-on-Latency (LoL), Low-on-Latency+ (LoL+), Bang-on-Bandwidth (BoB), Size-aware Rate Adaptation (SARA), Content-aware Playback Speed Control (CAPSC), Head-motion-aware Viewport Margins (HMAVM).We evaluate the proposed new algorithms using the objective metrics discussed in detail and show significant contributions for these new algorithms including up to 91% decrease in rebuffering duration for on-demand streaming, 61.9% decrease in rebuffering duration and 8.1% decrease in latency compared to L2A for low-latency live streaming, 81.3% bandwidth prediction accuracy for interactive streaming, lastly 20% improvement in viewport quality and 50% reduction in motion-to-high-quality delay for 360-degree video streaming.
A self establishing clustered network architecture for blockchain
Blockchain technology has branched out into many industries, such as healthcare, manufacturing, agriculture and entertainment, in the shape of both of its public and non-public variants. In principle, blockchain provides these industries with an immutable ledger, allowing the processes in its application environment to be taken care of in a decentralized manner. However, some challenges blockchain has faced to this day remain, such as the degree of its scalability, the level of security it provides and the transparency of the network transactions. In this thesis, a novel approach to a distributed, permission-less blockchain network is explored with the use of hierarchical clustering to gather the nodes based on the latency of their connection to one another. These clusters of nodes are allowed to work on their respective local chains and to add the verified local chains to the actual global chain that is used by the entire system. Network's throughput performance and overall latency are evaluated and compared with other blockchain applications, namely a simulation of the Bitcoin network itself and another approach that makes use of a method called Community Clustering. We collected the data for the correlation in the same environment for our work, Bitcoin and Community Clustering\cite{communityclustering} networks. The comparison of the collected data aligns with our work's clusters to improve the transaction throughput of the network, where an increase in average throughput and a drastic decrease in latency are observed.
Towards transparent recommenders: An explanation-based negotiation approach
As more and more recommendation systems are used in different areas and they are exposed to more ethical concerns, there is a growing demand for transparent and persuasive interactions with these systems. Toward this end, incorporating explainability in recommendation systems has emerged as a promising approach to enhance sociability and user trust. This thesis focuses on recommendation systems that utilize explainability techniques to foster sociability by providing precise and understandable explanations for their recommendations. The proposed recommendation system utilizes a combination of data-driven transparent mechanisms and human-agent negotiation approaches. The system generates personalized recommendations based on individual preferences and other similar user-tailored factors and engages in a negotiation with the users via discussions through explanations and real-time feedback mechanisms. The system reacts to user responses online, tailoring subsequent recommendations and explanations to convince the user. This thesis encompasses Nutrition Virtual Coach (NVC) agents that generate personalized food recommendations based on individual factors like allergies, eating habits, lifestyles, and ingredient preferences. It mainly focuses on explanation generation techniques to enhance the transparency and trustworthiness of the system by improving the NVC agent's sociability in multiple steps. Ultimately, we incrementally conducted multiple experiments with participants from various backgrounds to evaluate the acceptability and effectiveness of the system. The findings from the experiments generally indicate that most participants appreciate the opportunity to provide feedback and receive explanations for the given recommendations. The participants prefer receiving information tailored to their specific needs and expectations. Additionally, the participants expressed their thoughts on various forms of explanations. The findings indicate that comparative explanations are not appreciated as much as informative explanations. The users seem to prefer direct and simple explanations that explain items respectively.
Test betiklerindeki kod klonlarının otomatik tanımlanması ve tekrarın ortadan kaldırılması için yeniden yapılandırılması
Code clones in test scripts can significantly reduce maintainability. They can lead to duplication especially in ecosystems and product families where applications have commonalities in their test scenarios. Our goal is to eliminate duplication in test scripts for reducing their size and improving their maintainability. In particular, we focus on Python test scripts and propose an automated approach to identify commonalities, detect variations, and apply refactoring. We compare the effectiveness of various clone detection tools in identifying code clones at the function level in Python test scripts. We apply static analysis to detect variations and code dependencies among similar functions to steer the automated refactoring process. The refactoring approach involves extracting new functions or generalizing existing functions through parameterization. We validate our approach through case studies on one industrial and three open-source projects. Our findings indicate that automated refactoring can lead to a reduction of up to 8% in the size of test scripts. Automated detection and refactoring of code clones in test scripts can eliminate code duplication and reduce the script size. The proposed approach, validated through case studies, shows promising results and can be applied in various software ecosystems to improve test script maintainability
Yazılım kütüphanelerinin gruplandırılması
Most existing software modularization approaches focus on balancing cohesion and coupling to cluster software modules effectively. However, the criteria for modularizing software libraries should be addressed differently, given their unique characteristics. In these systems, cohesion emerges as the primary criterion, defined by the common usage patterns of modules rather than their interdependencies. Modules are grouped not because of strong coupling with each other but because they are often used together by other modules. We propose a novel optimization model based on this perspective and evaluate its effectiveness with two case studies. The first study involves an industrial application from the consumer electronics domain. The second study examines the transition from Java 8 to the Java 9 Platform Module System, where existing packages are reorganized into a higher-level unit of packaging and encapsulation, called modules. Our evaluation focuses on the efficiency of clustering, measured by the number of clusters required by modules and the reduction in redundant module imports. By grouping modules into clusters, importing a few clusters makes all its modules available, reducing overall imports. However, the goal is to balance this reduction by ensuring that only relevant modules are included in each cluster to prevent redundancy. We compare the results of our model against manually created clusters and alternative solutions generated by existing clustering algorithms, demonstrating significant performance improvements.
IoT için blok zinciri tabanlı ve enerji farkındalıklı uyarlanabilir güven mekanizmaları
Modern Internet of Things (IoT) increasingly rely on distributed decision-making, computation, and communication between independent nodes. As these systems expand and diversify, particularly when central coordination is not possible or not acceptable, decentralization becomes the primary concept. Decentralized systems eliminate dependence on central authority or trusted third parties by distributing decision-making and verification tasks among independent members of a network. While this shift promises greater scalability and fault tolerance, it introduces a critical challenge: establishing trust among resource-constrained and anonymous devices with no prior relationship or shared context. Blockchain technology has emerged to create a fundamental infrastructure to address this challenge by enabling transparent, auditable, and tamper-resistant management of data and consensus management in decentralized environments. Its distributed ledger mechanism ensures that transactions and state changes are immutably recorded, ensuring accountability in trustless networks. However, the presence of blockchain alone does not guarantee trustworthiness, especially when consensus mechanisms fail to account for the behavior, honesty, or sustainability of participating nodes. In this context, trust management becomes essential to maintaining the security and reliability of decentralized IoT networks. Trust can generally be broadly defined as the confidence in an entity's behavior based on past interactions, feedback, and reliability in service delivery. Modern trust models, ranging from reputation-based to computational and voting-based schemes, attempt to quantify this notion, yet most fall short in dynamic and malicious conditions. Despite extensive research, many existing trust models rely on static scores, stake-based voting, or unweighted historical data, making them vulnerable to adaptive adversaries and resource exhaustion. Especially for large-scale and heterogeneous IoT networks, lack of adaptive and behavior-aware trust evaluation diminishes the overall security and reliability. To address these limitations, this thesis introduces a series of progressively enhanced trust-aware models, each developed based on the insights gained from its predecessors: TrustVista introduces a layered trust evaluation model that calculates trust at the node, group, and community perspectives. It emphasizes recent behavior by using a weighted average that gives more importance to recent interactions, and applies statistical techniques to detect abnormal or potentially malicious behavior. By evaluating trust not just individually, but in a collaborative and structural context, TrustVista improves the speed and accuracy of identifying harmful participants. During this stage, it was observed that mechanisms like Delegated Proof of Stake (DPoS) could be a viable consensus mechanism for IoT, yet lacked any awareness of nodes' energy capacity. Based on this insight, EDPoS was developed to extend classical DPoS by incorporating energy-awareness. EDPoS enhances the classical DPoS mechanism by introducing energy-awareness into the consensus process. It dynamically adjusts how many nodes can become validators, how much data is included in each block, and which nodes are eligible to participate, based on how much energy nodes have and how efficiently they've used it in the past. This reduces the likelihood of block validation failures due to node dropouts and extends the overall system lifetime in IoT environments where resources are limited. Building on EDPoS, TEDPoS further integrates behavioral trust into the energy-aware DPoS framework. TEDPoS proposes a dynamic validator scoring method that combines three key factors: a node's trust level based on its behavior, its available energy, and the level of community support it receives through voting. Unlike traditional DPoS, TEDPoS adapts its selection process based on real-time network conditions, favoring nodes that are both trustworthy and resource-sufficient. This enables robust, context-aware, and attack-resilient delegate selection in decentralized IoT environments. As a supplementary use-case, TEMBUV, presented in Appendix~\ref{app:tembuv}, presents how trust and blockchain can jointly support secure operations in post-disaster communication networks. While not simulated in OMNeT++, this model offers a mathematical abstraction of trust-based decision-making in UAV and VANET-integrated domains. All proposed models are implemented and and tested against various adversarial conditions, including camouflage, bad-mouthing, and self-promotion attacks. Experimental results demonstrate clear improvements in malicious node detection, energy efficiency, and consensus reliability compared to baseline models. Collectively, the frameworks presented in this thesis contribute to the development of adaptive, secure, and scalable trust mechanisms for decentralized and heterogeneous IoT systems.
Protein-protein etkileşim tahmininde negatif örnekleme stratejilerinin karşılaştırmalı analizi
The accurate prediction of protein-protein interactions (PPIs) between host and pathogen proteins is essential for understanding viral infection mechanisms.However, a significant challenge is the lack of experimentally verified negative interactions (i.e., non-interacting pairs), which makes the development of effective negative sampling strategies essential. This thesis presents a comparative analysis of negative sampling strategies, specifically cluster-based methods. It proposes a novel Clustering-based Negative Sampling (CNS) method that aims to minimize the incorrect selection of negative interactions as positive interactions (i.e., experimentally confirmed interacting pairs). Furthermore, a structured evaluation methodology, Reliability of Sampling to Avoid False Examples (R-SAFE), is introduced to quantify the accuracy of these negative sampling methods. Experimental evaluations using four publicly available virus-host interaction datasets show that the characteristics of the datasets play a crucial role in the reliability of the negative sampling strategies. Therefore, a decision tree model is used to recommend the most appropriate negative sampling strategy based on the dataset features. The results emphasize the importance of negative sampling strategies to improve the predictive reliability of PPI models.
Sürekli entegrasyonda hata tahmini için bir makine öğrenmesi yaklaşımı
Continuous Integration (CI) is a development practice where developers regularly merge their code changes into a central repository, enabling simultaneous collaboration across a shared codebase. This frequent integration and automated building process in CI helps to detect and resolve conflicts or errors early in development. However, in large-scale systems, the build process can be costly. Each build incurs expenses, while skipping builds can increase the risk of undetected failures. This paper presents an empirical study within an industrial setting, investigating the use of machine learning techniques to predict build failures. Accurate predictions can help to identify builds that can be safely skipped to reduce CI costs. We evaluate various models and feature combinations on a dataset derived from real-world industrial projects. We observe high precision but low recall in predicting failed builds, allowing hundreds of successful builds to be correctly skipped, with around a dozen failures potentially being missed.
Algılama ve haberleşme kısıtları altında çoklu-İHA sistemi ile işbirlikçi arama
Multi-Unmanned Aerial Vehicle (UAV) systems have many uses in civil applications including Search and Rescue (SaR), surveillance, and environmental monitoring. In dynamic missions such as SaR, where multiple targets or areas of interest might need to be searched and monitored, the accuracy of target detection and information sharing among UAVs and with Ground Control Station (GCS) play a vital role in whether the mission will be successful or not. The state of the art sensors are still imperfect, hence, UAVs might need to take multiple measurement in each area of interest to increase the accuracy of sensor measurements. To that end, in this thesis we propose a joint area coverage, connectivity, and revisit time optimization framework within the context of multi-UAV multi-target cooperative search missions involving dynamic information exchange under imperfect sensing and communication constraints towards the goal of shortening time elapsed between consecutive visits to be able to take multiple sensor measurements as fast as possible (without suffering great losses in performance in terms of area coverage and connectivity), therefore increasing the accuracy of sensor measurements in the early stages of the mission in question. We utilize the Non-Dominated Sorting Genetic Algorithm (NSGA)-II evolutionary algorithm to solve the proposed optimization problem and compare our results on various scenarios. Our results show that by considering cell revisit times as well as connectivity in the path planning in the context of evolutionary algorithms, we observe up to 4-fold improvement in terms of observed times between visits and we see that the target detection times can be improved by 60% and we can achieve instant GCS inform post-detection in realistic, uncertain environments. Keywords: search and rescue, area coverage, drone networks, unmanned aerial vehicles, path planning, maintaining connectivity, multi-visit, target detection
Gizlilik temelli ortak süzgeçleme yöntemlerinin başarımının iyileştirilmesi
Privacy-preserving collaborative filtering methods offer useful filtering skills without deeply jeopardizing individual privacy. However, they mostly suffer from accuracy, scalability, and sparseness problems. Applying privacy measures to conceal confidential data in recommendation systems causes a bias in collected data, which might make accuracy worse. As the content in recommendation domain proliferates, the size of collected data expands rapidly, which aggravates scalability challenge of those systems. In addition, since users are typically able to rate a small fraction of existing products, sparseness of collected data becomes an issue.In this dissertation, various preprocessing methods are proposed to overcome accuracy, scalability, and sparseness challenges faced by various privacy-preserving collaborative filtering systems. Through application of the proposed preprocessing techniques like item ordering and elimination, clustering, dimensionality reduction, user profiling, profile cloning, and son on, novel privacy-preserving collaborative filtering schemes are cultivated. Essentially, the proposed enhanced systems focus on producing accurate predictions while coping with constantly growing nature of collections without jeopardizing individual privacy. The proposed schemes are analyzed in terms of privacy and overhead costs. Also, real data-based experiments are performed to scrutinize their effects on accuracy, scalability, and privacy. The analysis and experimental outcomes demonstrate that the methods preserve individual privacy and offer adequately accurate recommendations in scalable amount of time.Keywords: Preprocessing; Privacy; Scalability; Accuracy; Sparsity; Collaborative ?ltering.
Küme örtüsü probleminin genetik algoritma ile çözümü
Küme Örtüsü Problemi (KÖP), çeşitli alanlarda uygulamaları olan bir kombinatoryal optimizasyon problemidir. Seçilen kümelerin toplam sayısını en aza indirirken tüm öğeleri kapsayacak şekilde belirli bir uzaydan alt küme seçmeyi amaçlar. Doğal evrim süreçlerinden esinlenen Genetik Algoritmalar (GA), karmaşık optimizasyon problemlerinin çözümünde umut vaat eden sonuçlar vermektedir. Bu çalışmada, KÖP'ün üstesinden gelmek için başlangıç popülasyonunda kümelerin frekanslarına göre sıralama yapan bir sezgisel algoritma kullanılarak genetik algoritma ile çözüm sunulmaktadır. Ayrıca çözüm kümesine iyi katkı yapacak alt kümelerin seçimlerine öncelik verilmesi için bir seçim formülü geliştirilmiştir. Önerilen algoritma, seçim, çaprazlama ve mutasyon gibi genetik operatörleri kullanarak bir çözüm popülasyonunu geliştirmeyi amaçlamaktadır. Bireylerin uygunluğu, minimum sayıda küme ile tüm elemanların örtülme yeteneklerine göre belirlenmektedir. Genetik algoritma, nesiller boyunca çözümleri yinelemeli olarak iyileştirir ve kademeli olarak optimum veya optimuma yakın çözümlere yaklaşmaktadır. KÖP'ün çeşitli örnek problemleri üzerinde deneyler yaparak önerilen algoritmanın sonuçları gösterilmiştir.
Sosyal medyada sanal zorbalığın tespiti
Siber zorbalık, tüm dünyada olduğu gibi Türkiye'de de büyüyen bir sorundur. Şimdiye kadar elde edilen bulgulara göre, Türkiye'de sosyal medya kullananların siber zorbalığa maruz kalma olasılığı %20'i aşmıştır. Siber zorbalık tespiti İngilizcede çok olmasına rağmen Azerbaycan dili ve Türkçede çok az araştırma bulunmaktadır. Bu sorunu ortadan kaldırmak ve tespit etmek için genellikle makine öğrenimi kullanılmaktadır. Bu çalışmamızda, Azerbaycan dili ve Türkçe metinler üzerinde yapılmış siber zorbalıkları tespit etmek için farklı makine öğrenmesi algoritmaları kullanılmıştır. Çalışmamız, toplam 4400 adet Azerbaycan dili ve Türkçe yazılmış ve sosyal medyadan toplanan cümlelerden oluşan bir veri seti üzerinde makine öğrenimi teknikleri kullanılarak yapılmıştır. Sınıflandırıcıların performansını değerlendirmek için kesinlik (precision), doğruluk (accuracy), duyarlılık (recall) ve F1-skor kullanılmıştır. Çalışmada, kullanılan iki farklı veri setini de ele aldığımızda Türkçe veri setine göre CountVectorizer için %85.98 doğruluk ve %96.94 F1-skor ile Linear SVM modeli en yüksek sonuçlar vermiştir. Yine aynı model ve veri seti ile Tf-IdfVectorizer için en yüksek %85.77 doğruluk ve %97.85 F1-skor sonuçlarına ulaşılmıştır.
Yığılmaya dayalı ağ analizi
Ağlar, kimyasal sistemler, yapay sinir ağları, sosyal ağlar veya internet ve World Wide Web gibi farklı sistemleri modellemek için kullanılmaktadır. Bir iletişim ağı merkezlerden ve bu merkezleri birbirine bağlayan bağlantı hatlarından oluşur. Çizge kuramı bir iletişim ağının mimarisinin tasarımında ve analizinde güçlü bir matematiksel araçtır. Bir iletişim ağının temelini oluşturan topoloji bir çizge ile modellendiğinde çizgenin tepeler kümesi ağdaki merkezlere çizgenin ayrıtlar kümesi ağdaki bağlantı hatlarına karşılık gelir. Ağdaki önemli bir tepenin zarar görmesi çoğu durumda tüm ağın zarar görmesine yol açar. Bu nedenle önemli tepelerin belirlenmesi karmaşık ağların temel tepelerini koruyarak ağların sürdürülebilirliğini ve sağlamlığını arttırmaya yardımcı olur. Karmaşık ağlarda tepelerin öneminin belirlenmesi problemini çözmek için yığılmaya dayalı tepe büzülmesini temel alan bir yöntem önerilmiştir. Bu tezde çevre, yol ve benzeri çizgelerin yığılma değerleri hesaplanmış ve sonrasında büzülme işleminden yararlanılarak ilgili çizgelerin tepelerinin önem değerleri hesaplanmıştır. Bir çizgenin tepelerinin önem değerlerini bulan polinom zamanlı bir algoritma önerilmiştir.
Yapısal olmayan metinler için adlandırılmış varlık tanıma algoritmaları ve uygulamaları
Adlandırılmış varlık tanıma (AVT) problemi, veri çıkarımı, doğal dil işleme ve metin madenciliği gibi alanların alt dalı olarak ele alınmaktadır. Adlandırılmış varlık tanıma, yapılandırılmamış metinlerdeki varlık isimlerinin uygunluklarına göre önceden belirlenen kişi ismi, organizasyon ismi veya yer ismi gibi sınıflara atama yapmak için kullanılan bir araçtır. AVT çalışmaları pek çok alanda kullanıma sahiptir. Bunlara örnek olarak sohbet botlarının oluşturulması, sosyal ağlarda içerik önerisi oluşturma, özgeçmişleri işlemek veya müşteri çağrılarını sınıflandırmak ve onlardan öngörü elde etmek vb. söylenebilir. Bu tez çalışmasında ilk olarak iki farklı durum üzerinde AVT yapılmıştır. İlk olarak İngilizce haber yazılarından oluşan bir veri seti üzerinde iki farklı ön eğitimli kütüphane olan Spacy ve Stanford NLP kütüphaneleri kullanılarak kişi adı, yer adı, organizasyon adı vb. varlık adları tanınmaya çalışılmıştır. Bu çalışmanın sonunda kütüphaneler ile elde edilen doğruluk oranları, kütüphanelerin çalışma yapısı, hızları vb. ölçütler karşılaştırılmıştır. Çalışmanın devamında ise Twitter'daki Türkçe tweetler kullanılarak küfür, hakaret ve uygunsuz kelimeler adlandırılmış varlık tanım problemi olarak ele alınmış ve bu kelimeler farklı yöntemler ile tespit edilmeye çalışılmıştır. Önce metinlerde geçen kelime ve kelime öbekleri etiketlenmiş daha sonra ise etiketlenen kelimeler vektörleştirilmiştir. Vektörler, RNN, çift yönlü RNN, GRU, çift yönlü GRU, LSTM, çift yönlü LSTM ve önceden eğitilmiş çok dilli BERT modeli kullanılarak eğitim yapılmıştır. Modellerin çalışma sonuçları analiz edilmiş ve iki modelin sonuçları kıyaslamalı olarak değerlendirilmiştir.
Bulanık mantık tekniği kullanılarak stres değerlendirme
Individuals face many events during their lives. The human body reacts to all kinds of events that the individual encounters. This nonspecific response of the body to unexpected conditions and any change is called stress. Since human's perspectives on events are not the same, their reactions may not be the same. Therefore, the stress levels of individuals can be different. In order to minimize the damage caused by stress, which is an inseparable part of life, the first step is to determine the stress level of the person. Biomedical methods and scales are frequently used to determine the stress level. However, thanks to the development of technology, methods of determining the stress level have spread to fields such as artificial intelligence, machine learning, image processing, and neural network. In this thesis, a model based on fuzzy logic techniques has been developed using photographs containing facial expressions in order to assess the stress levels of individuals. Basically, the difference between two photographs of the individual was used in the model. This difference was obtained from eye pair, mouth, forehead and eyebrow regions, which contain the important symptoms of stress on the human face and were detected using image processing. Fuzzy logic techniques used in the application are Fuzzy C-Means (FCM) clustering, Adaptive Neuro-Fuzzy Inference System (ANFIS) and Fuzzy Inference System (FIS). In the final result obtained from the model was determined so that the stress level of the individuals was included in one of the "None", "Low", "Moderate" and "High" levels.
Mermer kalitesi sınıflandırması için derin öğrenme destekli bir model
The basic policy of marble enterprises is to establish sustainable high-quality products in a standardized manner. Identification and classification of different types of marbles is a critical task that is usually carried out by human experts. However, marble quality classification by human experts can be time-consuming, error-prone, unreliable, and subjective. Automated and computerized methods are needed to obtain more reliable, faster, and less subjective results. In this study, a deep learning model is developed in order to perform multi-classification of marble slab images with six different quality types. Some special image pre-processing operations were applied to the images for data augmentation and a special convolutional neural network (CNN) architecture was designed and implemented. It has been observed that the data augmentation approach for marble image samples has significantly improved the accuracy of the CNN model. Some outstanding results have been obtained with the proposed CNN model, which surpassed the alternative machine learning algorithms and even equalized the human experts' classification performance.
Deri hastalıklarının risk değerlendirmesi için görüntü işlemede bulanık teknikler
Many skin diseases are now documented in the literature. There may be some difficulties in diagnosing them while it is simple to observe them. Mole is a pigmented spot or small permanent protuberance on the human body that can be divided into malignant and benign categories. Melanoma is the medical term for a malignant mole. Melanoma-induced skin cancer is one of these observable diseases. Early diagnosis is critical for this cancer, which is relatively easy to treat once discovered. But disease diagnosis is a difficult stage in medicine. Diagnosis is based on a patient's signs, symptoms, physical examination, and some tests. However, a single human experience or opportunities may not be enough for the correct diagnosis. Such situations cause problems when doctors make the diagnosis. Developing computer technologies and machine learning techniques are applied to systems that will help doctors in the early diagnosis of diseases. In this study, a decision support system is studied for the detection of melanoma.
Çizgelerin diferansiyelinin hesaplanması
G = (V (G),E(G)) çizgesinin, her D ⊆ V (G) kümesi için, B(D), D tepe kümesinde komşusu olan V (G) \ D kümesindeki tepelerin kümesi olsun. D kümesinin diferansiyeli ∂(D) = |B(D)| − |D| ve G çizgesinin diferansiyeli ∂(G) = max{∂(D) : D ⊆ V (G)} olarak tanımlanır. ∂(D) = ∂(G) eşitliğini sağlayan D kümesine ∂-kümesi veya diferansiyel küme denir. Bir çizgenin diferansiyelinin araştırma ve uygulama alanı temel olarak hesaplamalı matematiktir. G çizgesinde D tepeler kümesi için eğer V (G) \ D kümesindeki her tepe D kümesindeki bir tepeye komşu ise D kümesine baskın küme denir. G çizgesi, aynı zamanda baskın küme olan bir ∂-kümesi içeriyor ise G çizgesine baskın diferansiyel çizge denir. Bu tezde ilk olarak, yol, çevre ve tekerlek benzeri çizgelerin diferansiyel değerleri hesaplanmış ve baskın diferansiyel olan çizgeler belirlenmiştir. Daha sonra, temel çizgelere ait tümleyen prizmaların diferansiyel değerleri hesaplanmış ve baskın diferansiyel olan tümleyen prizmalar belirlenmiştir. Bir çizgeye ait parametrelerle ilişkili olarak, bu çizgenin tümleyen prizmasının diferansiyel değeri araştırılmıştır. Son olarak, bir çizgenin diferansiyelini hesaplayan algoritma önerilmiştir.
Pamuk yetiştiriciliğindeki hastalıkların derin öğrenme yaklaşımı ile tahmin edilmesi
In this thesis, a study on the detection and prediction of cotton diseases, which is a sub-title of environmental factors that are effective in the cultivation of cotton plants, with the help of image processing and deep learning methods is presented. In the first stage, the images of the cotton plant were preprocessed in order to minimize the problems that may be encountered during the application of the preferred deep learning methods. These data obtained as a result of the preprocessing were used as input data for the optimization of the applied deep learning models. With the help of this input data, the hyper-parameters of Convolutional Neural Networks, Long Short-Term Memory Networks and Convolutional Long Short-Term Memory Networks models are decided. In the last phase, the success rates of the predictions made on random images given as input to these optimized models were evaluated. The results obtained as a result of the study were analyzed and compared with the studies in the literature.