27
Arşivlenen Tez
0
DOI Atanmış
0%
DOI Oranı
Anabilim Dalı
Optimization models to identify key RNA regulatory modules in cancer
Micro RNAs (miRNAs) are known as the important components of RNA silencing and post-transcriptional gene regulation, and they interact with messenger RNAs (mRNAs) either by degradation or by translational repression. miRNA alterations have a significant impact on the formation and progression of human cancers. Additionally, finding primary tissue of tumors is a fundamental property of a precise treatment. Therefore, employing computational methods with the ability to identify both tissue-specific and cohort-specific miRNA-mRNA regulatory modules is fast becoming a crucial tool in cancer biology. In this thesis, we first identified regulatory modules of 32 cancer types using a sparse multivariate factor regression (SMFR) model on matched miRNA and mRNA expression profiles of more than 9,000 primary tumors. We used an algorithm that decomposes the coefficient matrix into two low-rank matrices with separate sparsity-inducing penalty terms on each. Although, our solution significantly outperformed another decomposition-based approach in terms of normalized root mean squared error, the interpretability of the algorithm was not satisfying, due to the low sparsity level of solutions. Therefore, we proposed a single task two-step framework to model miRNA-mRNA relationships and identify cancer-specific modules between miRNA and mRNA from their matched expression profiles. We first estimated the regulatory matrix between miRNA and mRNA expression profiles by solving multiple linear programming problems. We then formulated a unified regularized factor regression (RFR) model that simultaneously estimates the effective number of modules (i.e., latent factors) and extracts modules by decomposing regulatory matrix into two low-rank matrices. Our RFR model groups correlated miRNAs together and correlated mRNAs together, and controls sparsity levels of both matrices. These attributes lead to interpretable results with high predictive performance. To find the biological relevance of our approach, we performed functional gene set enrichment and survival analyses. A large portion of the identified modules are significantly enriched in Hallmark, PID and KEGG pathways/gene sets. To validate the identified modules, we also performed literature validation as well as validation using experimentally supported miRTarBase database. By applying proposed single-task algorithm on 32 independent cancer cohorts, interestingly we observed similar patterns in identified miRNAs and mRNAs, also in significance of certain gene sets for specific cohorts. Inspired with the previous results, we attempted to investigate the similarities between mechanism of miRNA-mRNA regulatory modules in different cancer cohorts from the same tissue to improve the predictive performance of the model and extract more detailed and informative solutions. Moreover, detecting commonalities of cohorts from the same tissue, can provide valuable information about the underlying tissue and cohort mechanisms. Hence, we established a multitask learning formulation to identify key tissue- and cohort-specific miRNA-mRNA regulatory modules from their matched expression profiles of tumors. To this end, we proposed a multitask learning sparse regularized factor regression (MSRFR) method to model the sparse relationship between miRNAs and mRNAs, extract tissue- and cohort-specific miRNA-mRNA regulatory modules separately and estimate the number of cohort-specific regulatory modules and shared modules (i.e., tissue-specific regulatory modules). As the validation process, we first performed literature survey to find the relationship of identified cohort-specific miRNAs and their corresponding cancer. We then applied experimentally supported validation of miRNA-mRNA interactions using miRTarBase database. Finally, to validate the performance of MSRFR model in distinguishing tissue- and cohort-specific regulatory modules, we explored the biological mechanism of identified key regulatory modules by checking their enriched pathways in PID and Hallmark collections. The overall result indicated that proposed model was able to effectively extract meaningful miRNA-mRNA regulatory modules and differentiate tissue- and cohort-specific modules.
Developing data-driven methods using machine learning in operations and finance
With the recent advances in data collection and analysis technologies, data-driven approaches to operations management and financial problems have gained traction. In particular, machine learning methods are increasingly being integrated into optimization problems. This integration facilitates translating the historical data into prescriptive solutions in comprehensive frameworks. This thesis aims at building data-driven methods to link the data with the actions and improve the goals in operational and financial decision systems. We consider an integrated learning and optimization approach using machine learning techniques for optimizing strategy for decision-makers facing a complex problem with additional information about the state of the system. We give algorithms based on integrating optimization, neural networks, and adapting time series properties, and develop models that are capable of estimating nonlinear relations between data. We focus on three problems from the fields of inventory, financial investment, and firms evaluation. In the context of inventory, we consider an integrated learning and optimization problem for optimizing a newsvendor's strategy facing a complex correlated demand with additional information about the unobservable state of the system. We, therefore, combine estimation, inference, and optimization using a multi-layered neural network. To assess the performance of this integrated approach, we compare the results from our approach against data-based methods that ignore the hidden factor information or that employ separate inference and optimization steps. Numerical examples on both a synthetic data set and real data which might have an unobservable state demonstrate that our approach compares favorably against the other benchmarks. In the context of financial investment, we propose a new approach to asset allocation based on machine learning; it analyzes historical market states and asset returns and identifies the optimal portfolio choice in a new period when new observations become available. In this approach, we directly relate state variables to portfolio weights, rather than first modeling the return distribution and subsequently estimating the portfolio choice. The method captures nonlinearity among the state (predicting) variables and portfolio weights without assuming any particular distribution of returns and other data, without fitting a model with a fixed number of predicting variables to data, and without estimating any parameters. The empirical results for a portfolio of stock and bond indices show the proposed approach generates a more efficient outcome compared to traditional methods and is robust in using different objective functions across different sample periods. In the last project, we propose a nonlinear approach based on stochastic frontier analysis and machine learning to estimate the pricing efficiency and the level of premarket inefficiencies for initial public offerings (IPOs). This approach enables us to estimate IPO pricing efficiency using information available before the IPO day and without any distributional assumptions for deliberate underpricing in the premarket that have been documented in the literature. We apply the proposed approach in the U.S. IPO market and show that only a few determinants of the value of firms impact the pricing and underpricing of IPOs.
MDP model for the preference-based appointment scheduling problem with multi-priority patients
This thesis consists of three parts that analyze various aspects of appointment scheduling procedures in healthcare facilities. In the first part, we study an outpatient clinic with multiple types of patients that contact with the clinic to schedule an appointment. The clinic dynamically decides on the set of appointment days offered to each patient. Patients have different utility weights for each day in the booking horizon. Based on these utility weights, they either select one of the appointment days offered in the set or they leave the system without making an appointment. We model this system using a periodically time-inhomogeneous Markov decision process and derive analytical results on the structure of the optimal policy. To develop a solution for the model, we implement a simulation-based booking limit improvement algorithm by approximating the value function. In the second part, we consider a primary care clinic with strategic patients who choose between making an appointment, with an indirect wait cost, and walking in, with an inconvenience cost and a risk of being rejected. We consider two types of patients, regular, and urgent, based on their indirect waiting cost. Considering the equilibrium behavior of patients, the clinic should determine the optimal number of slots reserved for walk-ins, to maximize expected revenues. We analyze the system under observable and unobservable indirect waiting time information and characterize the equilibrium patient behavior for both of the settings. Finally, in the third part, we consider a model that combines patients' walk-in, no-show and cancellation behavior and appointment day preferences. In this study, the clinic decides on the number of slots allocated for walk-in patients and the set of appointment days offered to a patient. Given the offered set, patients may select making an appointment for one of the offered days or walking in on a day that is not included in the set. Walk-in patients have a risk of not receiving service, which we call "blockage" and blockage probability of a patient depends on the other patients' walk-in behavior. Thus, patients consider the other patients' choices while making a walk-in decision, which results in endogenously determined walk-in rate. We formulate the problem with a deterministic fluid model to maximize the expected net profit. Furthermore, we characterize the structure of the optimal fluid solution and establish its asymptotic optimality.
Multi-objective optimization in cement and textile industries using the triple-bottom-line accounting for sustainable production
With the increasing environmental disasters and growing social awareness sustainability has become the cornerstone/key to the operation and design of supply chain management for organizations and governments. The triple bottom line accounting is a comprehensive tool for researchers and practitioners in incorporating the three pillars of sustainability i-e Economy, Environment, and Society. So far research has been more focused on economic objectives and recently to large extent on environmental pillars. Research on all three pillars simultaneously is still lagging behind. Therefore, in order to analyze the complete picture of sustainability a more comprehensive approach is to be applied. The TBL is a methodology that takes all three dimension of sustainability at the same time for a sustainable decision-making process. In this thesis, a systematic approach is used to incorporate sustainability considerations in resource consuming industries of textile and cement production. The proposed framework is based on identifying sustainability indicators in textile and cement industries. Then validating these indicators if they are not previously validated in literature and practice, developing a MOMILP and then generating Pareto optimal solutions. We have used the 3 S methods for validation. This method use the self, scientific and society (hence the 3S) for indicator validation. Next, we formulated a generalized MILP for both Textile and Cement Industries incorporating our sustainability indicators. Lastly, we have used two different solution strategies for multi-objective optimization problems,(i) we have used AUGMECON2 for solving the MOMILP of textile industries and generated a Pareto frontier for balanced decision making. We performed a sensitivity analysis of the model to give further insights into the resultant solutions. (ii) For cement industry we have used a recently proposed multi-objective optimization method GoNDEFF. We have generated a set non-dominated solution points and efficient binary solution set.
Strategic customer behavior in service systems: Externalities, risk sensitivity and heterogeneity
This thesis considers the analysis of models of service systems where customers are strategic risk-averse and decide whether to join the service system or balk, by taking into account various uncertainties and externalities generated by other potential customers. First, we propose a game-theoretic framework for a static service system with strategic risk-averse customers where we ignore the dynamics of the service system such as random arrivals or random service times and seek capturing the equilibrium joining probabilities in the proposed framework by utilizing other attributes of the service system such as reward, externalities and risk-sensitivity. Next, the pricing problem is presented and analyzed. We then, present an M/M/1 queueing model where similar to the previous static framework, customers are risk-averse and strategic, and address the equilibrium joining behavior of customers in such a system. We also, come up with a pricing scheme for the service provider and social welfare maximizer. Interestingly, our results indicate the possibility of charging negative prices to achieve socially desired arrival rates with risk-averse strategic customers. Finally, we extend our results to the case where customers are heterogeneous in their risk-sensitivity degree in an M/M/1 model with strategic risk-averse customers.
New product introductions with unique and common features
Firms have to determine the right features and price for their new products as they introduce new ones to the market. While doing so, they are challenged to respond to the changes whether it is coming from their customers or competitors. In this dissertation, we focus on the problem of determining the optimal features of new products to be introduced into the market and its pricing as well as whether to discontinue the old product or keep it with the new one in case of a monopoly. We first focus on a monopolistic environment and consider the firm's problem of determining the optimal unique and common features for its new product given an existing product offered by the same firm, setting the optimal prices of both products, and finally deciding whether to still offer the old product alongside with the new one. We contribute to the literature by modeling a new product introduction problem of a monopolistic firm using discrete features and further differentiating between the unique and the common features to better capture the demand substitution effect. We then characterize and analytically solve the underlying model under linear cost, demand, and price functions to determine, in linear time, the optimal features of a new product as well as the optimal rollover strategy and the optimal prices for both the old and the new products of the firm. With our model, firms can express the product quality and price in terms of its unique and common features, and determine the optimal features for its new product without the necessity of going through all possible feature combinations. Our work in this chapter is also the first one to analyze the firm's rollover decision by using discrete product features. We then consider the problem of determining the features and prices of two competing firms' products in a duopoly where the firms introduce their products sequentially. The first firm introduces a product by determining the features and the price of its new product by anticipating the second firm's new product. The second firm then competes with both the product and price by introducing its new product after observing the first firm's product. We also comment on the case in which the second firm cannot change its offered product and consequently only responds with a pricing decision. We contribute to the literature with our model by capturing the product substitution effect on the feature quality level which also allows us to directly capture the effect of commonness and uniqueness of features on substitution effects. The advantage of our model is crucial in settings where the set of potential features and feature levels is finite and small such that it cannot be approximated by continuous variables. In the last part, we further explore product entry decisions with feature and quality level selections in monopoly and duopoly settings by extending the analysis from the second part. We first consider the simultaneous entry setting where the firms do not know the competitor's product choice beforehand and have to commit to their decision. Second, we consider a new product entry from one firm and a price-only response from the other firm subsequently over multiple periods. We then consider the setting where the second firm has a probability over its decision on whether to enter with a new product or respond with only a price. Lastly, we consider a monopolistic firm that tries to maximize its total profit by allocating features and quality levels to its two products having their own brands. The results in these settings complement our contributions from the previous part by showing that modeling products using unique and common features with varying quality levels is not only fully defining the products but also expandable to different settings that can capture the various market dynamics that can occur between the competitors and the competing products. In summary, we model and analyze the firms' selection of unique and common features for their new products in monopolistic and competitive market settings. We derive insights from analytical and numerical results that would help the managers to decide under these settings.
Proactive return management in E-commerce supply chains: Predictive analytics approach
In e-commerce, product returns management has become a critical concern for online retailers. With the drastic growth of online shopping, the increasing volume of returns poses significant challenges to the efficiency of supply chain operations. To effectively address these challenges, proactive return management strategies are essential. This MS thesis aims to predict product returns before the product is sold, leveraging open-source customer-item rating and feedback data. The study adopts a novel approach combining Natural Language Processing (NLP) techniques, Matrix Factorization based on Bayesian personalized ranking loss, and deep learning methodologies to uncover latent factors behind shopping and rating behavior to make accurate return predictions, aiding online retailers in proactive decision-making and optimizing their supply chain and inventory management processes.
Tramp shipment routing and scheduling with inventory and stowage constraints
Maritime transport has been a fundamental mode of transportation for centuries, enabling the transportation of heavy and bulky items in large quantities across oceans and seas. The use of ships has allowed for the efficient movement of goods and commodities and has been vital in facilitating global trade and commerce. In modern times, maritime transport has become even more critical with the exponential growth in global trade and the movement of goods, making it an essential component of the global economy. There are various types of operations involved in maritime transport, with liner, tramp, and industrial shipping being the three most commonly used. Liner shipping refers to the use of regularly scheduled services between designated ports, where ships follow a fixed timetable and transport a specific type of cargo. In contrast, tramp shipping operates on an ad hoc basis, with ships not following a fixed schedule but instead sailing according to the requirements of customers, and the ships are usually rented or leased. Industrial shipping, on the other hand, refers to the transportation of goods using ships that are owned by the supplier or customer. Tramp shipping alongside with the liner shipping is widely used in maritime transport, particularly for the transportation of bulk cargoes such as coal, iron, and grain. In this type of shipping, ships are chartered on a single voyage basis, and the cargo is loaded and transported according to the specific requirements of the customer. The ships used in tramp shipping are usually equipped with cranes to load and unload the cargo. In the tramp shipping problem considered in this study, the objective is to transport a set of cargoes from their origin ports to their respective destination ports using a homogeneous fleet of ships. The origin and destination ports of both the cargoes and the ships are predetermined. All of the cargoes are mandatory and have to be sent to their destination point within their respective deadline. If a cargo arrives at its destination point before the deadline, there will be a deduction from the cost as a reward. To minimize the sailing cost, inventory cost, and stowage cost, the problem is modeled as a mixed integer linear program, taking into account the stowage and inventory constraints that arise in real life applications. Stowage constraints are used to ensure the stability of the ship and to eliminate the loss of time when loading and unloading the cargoes. The ports have initial inventories of the cargoes, and their inventory value changes according to the amount shipped. Inventory constraints are developed to control the changes in inventory levels. Decisions must be made based on the routes and schedules of the ships, as well as the stowage and inventory constraints. The computational results obtained from 153 instances provides valuable insights for solving the tramp-shipping problem. The results show that while small-sized instances can be solved optimally, medium-sized and large-sized instances achieve gaps between 5-30% and 30-70%, respectively, depending on different problem characteristics. This gap is found by comparing the current solution to the best bound found in the problem. This study of the tramp shipping problem is designed to provide significant practical applications in the maritime transportation industry, as it can help shipping companies optimize their fleet utilization and reduce their transportation costs, leading to increased efficiency and profitability. Previous literature concerning the combination of these three maritime problems was limited, this study is designed to fill a significant gap in the literature and bring valuable insights for further research.
Advanced algorithms and solution techniques for U-shaped assembly line balancing problems
Assembly line balancing problems entail allocation of work content to different work stations, optimizing certain criteria like minimizing the number of stations, fair allocation of work content and cycle time minimization etc. The U type layout is known for its exceptional efficiency and flexibility for line balancing problems. It also provides workers with opportunities to enhance their skills and work experience through cooperation and continuous learning. However, most of the models in literature ignore the full potential of U-lines and focus on a few narrow aspects for line balancing. In this dissertation, we address the issue of assigning tasks to workstations in ways that offer more choices to line managers and workers. In the first part of this dissertation, we develop effective logic cuts that exploit the logical structure of the problems, to improve the performance of integer programming model in terms of computational time. Proposed logic cuts enable the model to utilize bin-packing bounds at each station. Moreover, idle time information, and the knowledge about combinations of task assignments to stations, that produce solutions dominated by alternative assignments, are also exploited. Computational experiments demonstrate that our enhanced model outperforms the previously developed integer programming models in solving U-type assembly line balancing problems In the second part, we present novel ways of assigning fair workload to stations. Our model is able to fulfill different fairness criteria like minimization of mean absolute deviation and sum of squared differences. We also develop a method for distributing certain types of workloads regularly along the assembly line. This way, work sharing and benefits of communication between the workers can be enhanced. Our approach also enables the decision makers to allocate appropriate idle time to certain groups of tasks. This third part deals with uncertain task times. We develop a new variance bounds based approach to deal with the stochastic U-line balancing problem. The variance based approach is considerably simpler to implement and solves the problems efficiently, outperforming other chance constrained models used for U-lines in literature. Overall, we present methods that can be adapted easily to more complex assembly line balancing problems and provide additional choices to managers, researchers and decision makers.
Efficient optimization algorithms for computational biology
The development of efficient optimization algorithms is crucial for computational biology due to the unique challenges and requirements of biological data. These algorithms may enable researchers to extract meaningful insights from vast and complex data sets, driving forward our understanding of biological systems and improving therapeutic interventions. In this thesis, we have developed algorithms for computational biology in cancer subtyping and drug-target interaction prediction. Identifying cancer subtypes is important for providing personalized treatment effectively, developing new drugs, characterizing risk factors, and understanding the underlying mechanisms of diseases. In the first part of this thesis, we present a clustering algorithm, named GSPS, that uses multiple kernels defined on pathways/gene sets for identifying cancer subtypes. GSPS employs an efficient decomposition algorithm for solving large scale optimization problems within the localized multiple kernel k-means clustering and provides a standalone framework for obtaining patient subtypes on cancer cohorts. We perform clustering experiments on gene expression profiles of primary tumors for 33 cancer types of the Cancer Genome Atlas using three different pathway/gene set collections. We compare our proposed method against three standard algorithms that can integrate pathway and gene expression profiles. Our approach shows statistically significantly better or comparable performance on survival analyses. Our method is also able to produce interpretable information between obtained cancer subtypes and pathway/gene set collections. In the second part of this thesis, we also propose a novel framework, manifold optimization based kernel preserving embedding (MOKPE), to efficiently solve the problem of modeling heterogeneous data. In many applications of bioinformatics, data stem from distinct heterogeneous sources. One of the well-known examples is the identification of drug-target interactions (DTIs), which is of significant importance in drug discovery and repurposing. Our model projects heterogeneous drug and target data into a unified embedding space by preserving drug-target interactions and drug-drug, target-target similarities simultaneously. We performed ten replications of ten-fold cross validation on four different drug-target interaction network data sets for predicting DTIs for previously unseen drugs. The classification evaluation metrics showed better or comparable performance compared to previous similarity-based state-of-the-art methods. We also evaluated MOKPE on predicting unknown DTIs of a given network. In this thesis, we also extended MOKPE, and developed MOKPE+, to use multiple drug-drug and target-target similarities with the aim of increasing the accuracy and interpretability of DTI predictions. For this purpose, using a localized approach, we followed a similarity selection and fusion method that has features such as estimating the similarity weights of previously unseen new drugs and cleaning noisy input. We performed ten-fold cross-validation with five replications to predict DTIs for new drugs on four different drug-target interaction network data sets. We used this similarity selection and integration method both with MOKPE+ and in the baseline models we have previously compared. We also used methods specifically developed to exploit multiple similarities. Classification evaluation metrics showed that MOKPE+ showed better or similar performance compared to both other baseline models and machine learning models that can use multiple similarities directly.
Integration of machine learning and optimization models for data-driven decisions: Applications to lot sizing problem with random yield
In recent years, significant advancements in data collection, analytical techniques, and data-driven methodologies have transformed the landscape of addressing complex decision-making and production planning challenges. Building on these innovations, this thesis introduces novel data-driven approaches to manage uncertainty and variability across diverse applications. The core focus of this research is to leverage big data in uncertain environments to derive effective decision-making rules for policymakers by integrating machine learning techniques with optimization methods. The thesis encompasses three primary studies: decision-making problems under uncertainty, a single-stage lot sizing problem, and a multi-period lot sizing problem with random yields. First, we address a challenge in decision-making and classification problems where actions must be taken prior to observing an uncertain event, and incorrect decisions result in asymmetric costs. We propose some novel methods that integrate machine learning and optimization techniques to derive decision rules from limited observations, particularly in scenarios where frequent features have a significant impact on outcomes. Our methods aim to minimize the costs associated with incorrect decisions in both binary and multi class classification problems. Experimental results using publicly available data sets show that our approaches significantly outperform traditional benchmarks, reducing costs by up to 95% with respect to naive benchmarks. Second, we investigate a data-driven lot sizing problem under random yield. Motivated by semi-conductor production, we focus on the case where the random yield rate of a manufacturing process depends on a large number of features that can be observed before the lot sizing decision is made. Similarly, demand may also be random and may depend on a number of features. The lot sizing problem in this setting is challenging because the optimal decision depends on a large number of observed features for which there is limited data. To address this challenge, we propose estimation and optimization methods that combine tools from machine learning with tools from stochastic optimization. Using a publicly available data set for semi-conductor yield data and an additional synthetic data set, we compare the performance of different estimation and optimization approaches. We show that there is significant value of taking feature information into account for cost minimization. We also find that the best method for this problem combines tools from estimation with theoretical optimization properties of the random yield inventory problem. Finally, we extend the single-period lot sizing problem to a multi-period setting. Given the realized variability in yields across periods, there are numerous potential scenarios by the end of the production horizon. The main objective is to develop a data-driven rule for determining optimal lot sizes under unknown yield distributions, unlike conventional approaches in the literature that assume known distributions. This rule aims to minimize the expected underage and overage costs at the end of the final period plus holding cost of intermediate periods and variable production cost while ensuring that total customer demand is met. This study advances the literature by introducing reinforcement learning to address the multi-period lot sizing problem with stochastic yields and unknown distribution for the first time. To assess the effectiveness of our approach, we compare results against established static and dynamic optimization methods. Our method provides a flexible and adaptive framework for decision-making in complex production systems characterized by stochastic yields.
Minimization of the spread of acute rumors in social networks
Social networks offer the capability to spread inspiring ideas and adapt innovations at unprecedented speed and convenience. Although such an information-spreading capability is invaluable, social networks can also rapidly disseminate misinformation to a large number of people, with dire consequences. In this study, we focus on the spread of a specific type of misinformation: acute rumors, which we define as misinformation with the potential to spread quickly in the network and mobilize individuals to take harmful actions that may go beyond social platforms. We present a new diffusion model for acute rumors, which extends the classical linear threshold model by considering the base idea of the individuals regarding the topic of the rumor and the emotional impact its content creates on them. Based on this diffusion model, we propose a new centrality measure that can accurately detect individuals with a high potential to enhance the spread of acute rumors. In a comprehensive numerical study, we evaluate the performance of our proposed centrality measure by comparing it to existing traditional centrality measures in the literature. Our results attest to the superior performance of our centrality measure not only in terms of detecting the individuals with the highest potential to increase the reach of acute rumor but also in finding the correct ranking among the individuals in the network regarding their potential to contribute significantly to the dissemination of acute rumors. While the proposed diffusion model and centrality measure offer valuable insights into who is likely to spread an acute rumor, they do not fully capture how platform-specific dynamics and individual-level behaviors shape the overall diffusion process. To address the limitations of the diffusion model in capturing message variability and algorithmically shaped user behavior, we develop an agent-based model (ABM) inspired by the structure of Twitter/X. The model is calibrated using empirical data and incorporates multiple forms of user engagement. Within this framework, we evaluate several control strategies intended to reduce the spread of acute rumors while maintaining user interaction on the platform and avoiding real-world physical actions triggered by rumor diffusion. Simulation results show that the targeted intervention, which adjusts the reading order for a selected subset of users with strong or polarized opinions identified through the proposed centrality measure, outperforms all other strategies and achieves the most effective balance between limiting rumor diffusion, preventing physical actions, and maintaining user interaction on the platform.
Dynamic relief provision planning for en route refugees in humanitarian logistics
The forced displacement crisis has become a pressing global humanitarian concern. Massive refugee movements force individuals into dire living conditions with severe inaccessibility to essential services and resources. Humanitarian organizations play a vital role in alleviating the hardships faced by refugees en route through relief aid interventions; however, the adoption of academic approaches to support these efforts remains limited. This thesis aims to optimize the periodic delivery of relief aid to geographically dispersed refugee groups en route to safe destinations, under both deterministic and stochastic migration settings. In the deterministic setting, we introduce a Capacitated Mobile Facility Location Problem with Mobile Demand and formulate it as a mixed-integer linear program (MILP). To solve this complex problem, we develop two solution approaches: an accelerated Benders decomposition algorithm for exact optimization and a matheuristic algorithm based on an enhanced fix-and-optimize strategy. Using realistic scenarios inspired by the Honduras migration crisis, our results show that the Benders decomposition reduces computation time by 46% on average while maintaining solution quality. The matheuristic, meanwhile, produces near-optimal solutions with only a 2.4% average optimality gap and significantly reduced computational effort. For the stochastic counterpart, we model refugee movements as a Markov decision process (MDP) with probabilistic transitions driven by migration pull factors such as safety, road conditions, and spatial proximity. To solve the MDP, we develop an approximate dynamic programming algorithm featuring customized basis functions and a novel policy replication mechanism. We further propose a state-dependent variable threshold policy that efficiently generates high-quality service plans. Applied to the Syria–Türkiye migration crisis, our approach achieves up to 25% cost savings over deterministic methods and up to 12% additional savings through coordinated planning across humanitarian organizations. The proposed methods are shown to be effective across diverse migration dynamics, including dispersed and cohesive refugee flows and multi-destination settings. Additionally, we identify service and traversal hotspots along migration routes, offering valuable guidance for tactical planning and resource pre-positioning. In summary, this thesis contributes cost-effective, scalable, and operationally relevant solutions for humanitarian aid delivery to en route refugees and provides actionable managerial insights for addressing future displacement crises.
The empirical analysis of international climate policies for energy decisions
While the Paris Agreement enforces a radical transformation of energy systems to limit global temperature rise to below 1.5°C, market-based economic instruments, like the European Union Emissions Trading System (ETS) and the Carbon Border Adjustment Mechanism (CBAM), accelerates the internalization of carbon emission costs internationally. This pressure demands transparent, feasible, and comparable planning tools to safeguard supply security while controlling costs. However, significant gaps persist: data-intensive general equilibrium models fall short in responsiveness, and machine learning-based prediction models, as "black boxes", lack sufficient auditability and transparency for policymakers. This dissertation, comprising a complementary series of four papers, aims to fill the identified gap by developing a multi-layered hybrid methodology building on the global energy analysis covering 147 countries and extending it to national dynamics of Türkiye and electricity sector in detail. This study integrates computationally efficient partial equilibrium economics, multi-output machine learning, and optimization methods to provide a controllable trade-off between cost, carbon, and energy security under ETS/CBAM conditions across various carbon price scenarios. It proposes an innovative planning perspective that simultaneously tackles the carbon constraints of the Paris Agreement along with the goals of energy security and fiscal efficiency, through four consecutively structured papers spanning global, national, and sectoral scales. In the first paper, the partial equilibrium model replicates the fossil fuel share of 147 countries from 1997–2014 with an average absolute error of 0.28–0.34%, demonstrating that carbon tax-incentive packages align with the 2°C target without disrupting budget balance. The second paper models the governance update period (GUP) using multilevel machine learning, quantitatively determining an optimal policy cycle that further reduces the fossil share by 6 percentage points. The third paper, scaled to the national level, assesses Türkiye's electricity installed capacity for 2030 by comparing a parametric depreciation model with LSTM/GRU-based deep learning estimates, where the deep learning approach significantly reduces the normalized error rate in hydropower generation. The Green Growth scenario increases the renewable energy share from 56% to 79% while lowering carbon intensity to 160 kg CO₂/MWh. In the fourth paper, multiple output (MO) predictions are integrated with set-covering and two-stage grid-search optimization, reducing the total root mean square error (RMSE) to 0.587, boosting renewable energy capacity by 2.2 times compared to 2024, and enabling the grid-search method to achieve the lowest CO₂ emission profile. The collective findings from these four papers underscore the development of a multi-layered hybrid methodology that delivers three unique contributions: (i) it facilitates seamless integration of global modeling outputs with low computational overhead into national-level machine learning models; (ii) it directly models simultaneous relationships among diverse energy sources through its multi-output architecture; (iii) it employs optimization tools in bidirectional interaction with machine learning forecasts, enabling a holistic evaluation of decision-making, cost, and risk dimensions within a unified framework. Findings indicate that robust carbon pricing and data-driven capacity allocation, paired with a long-term planning perspective, can align carbon intensity with Paris Agreement targets while minimizing total system costs. Consequently, methodology developed in this dissertation equips policymakers with a scalable, auditable, and feasible decision-support tool within the ETS–CBAM environment.
Kredi kartı işlemleri için üye işyeri hizmet ücreti optimizasyonu
This thesis aims to enhance the current pricing policies set to optimize customer retention by proposing enhanced methodologies, focusing on finding efficient algorithms that apply to large-scale businesses. Customer retention studies have a vital effect on companies since they affect the generated revenue and market share. This study mainly investigates pricing factors and their relationship with customer churn. The aim is to find commission rates that will be used to charge merchants when making transactions with Point of Sale (POS) devices. An important aspect to remember is the price sensitivity of those merchants as there happens to be a negative relationship between their retention, and pricing policies. We designed a nonlinear mathematical model to solve this problem. However, solving nonlinear models for large-scale problems is not effective when computational power and current solving techniques are considered. As an approach to solving this phenomenon, we discretized the nonlinear form of the model and proposed an alternative integer programming model. By doing so, solvers used to optimize the integer programming models were able to solve higher scales of the problem when compared with global nonlinear solvers. However, it is observed that IP solvers fail to solve the proposed mathematical model after a certain size. To solve this dimensionality problem, stratification techniques are used to divide the problem and optimize subparts of the whole problem separately while limitations and the objective of the problem are considered.
Türk dağıtıcıların karşılaştığı zengin araç rotalama problemi için metasezgisel algoritmaların ve çözüm süreçlerinin araştırılması
Logistical challenges pose a substantial financial burden for distributors worldwide, particularly in light of escalating oil prices and evolving customer demands. This thesis addresses a Rich Vehicle Routing Problem (RVRP) encountered by a distributor in Turkey. The problem involves a multi-objective function that integrates road tariffs, specific customer constraints, and features observed in various VRP variants. Our study explores diverse solving approaches, including an exact method developed through a novel Mixed-Integer Linear Program, as well as non-exact solutions employing heuristic and metaheuristic algorithms. Notably, our findings highlight the efficacy of metaheuristic algorithms, particularly through a two-phase solving strategy: initially dividing and solving the problem into individual sub-problems, then integrating them comprehensively. The study concluded with a 5.13% reduction in operational costs compared to the distributor's existing plan, indicating significant potential for enhanced profitability while optimizing the daily scheduling of the vehicle fleet, and further expanding applications in operations research.
Sehir lojistiğinde yer seçimi–rotalama ve senkronizasyon problemleri
City logistics aims to improve urban freight transportation by considering the costs and benefits of public and private sectors, consolidating segmented freight shipments, and integrating the individual actors in a collaborative environment. This thesis studies network design problems in city logistics systems to address managerial challenges in urban freight transportation. We consider two network design schemes, namely the one and the two-echelon distribution networks, to formulate strategic and tactical level problems in urban freight transportation. In the single-echelon systems, freight is distributed from consolidation centers located on city boundaries to the customers inside the city. In the two-echelon systems, goods are unloaded at the intermediate facilities, called satellites, and consolidated into smaller vehicles suitable for the last-mile delivery in city centers. From an operational level perspective, we highlight the importance of synchronizing first and second echelon vehicles at the satellite locations and discuss the relation of the satellite synchronization problem to the network design problems. We propose mathematical programming formulations for the introduced strategic, tactical, and operational level problems in city logistics, and develop exact and heuristic solution approaches. The exact approaches use column generation to find optimal delivery routes efficiently. The heuristics are based on the hierarchical decomposition of the original problem into its basic decisions. Extensive computational studies in this thesis provide new insights into designing and implementing a practical city logistics system in real-world.
Kuyruk tipi servis sistemlerinde kalite bilincindeki stratejik müşteriler
A service system is a configuration of technology and organizational networks designed to deliver services that satisfy needs, wants and aspirations of customers. Call centers, universities, hospitals, restaurants, shopping centers, hotels, beauty centers are a few examples of the service systems which provide service to customers in different areas. A strategic customer in a service system acts in order to maximize his individual welfare or utility. To maximize his individual utility, a strategic customer compares all of his possible alternatives and chooses the one which brings the highest pay-off to him. The alternative with the highest pay-off is referred as the optimal behavior (action) of this customer. Indeed, each customer's optimal behavior is affected by acts taken by the service provider (system manager) and by the other customers. The result is an aggregate equilibrium pattern of behavior which may not be optimal from the perspective of the society as a whole. So, the optimal decisions of the individual customer, social planner who is responsible from maximizing the social welfare, and the service provider all need to be analyzed in these systems. Many service systems are queueing type systems. So service systems generally include queues, and the individual customer, social planner and the service provider must (directly or indirectly) consider the displeasure of waiting in the queue. Moreover, the service systems that we consider in this thesis have imperfect quality. The service failures which are caused by the service quality problem will require resolutions. Designing the service systems by using different system structures we analyze different service failure resolution alternatives in Strategic Customer Setting. In summary, our goal in this thesis is to integrate service quality issues with strategic behavior and compare different service system structures
Afet sonrasında yolları açmak için çok-araçlı ayrıt rotalama problemleri
Doğal felaketlerden sonra köprüler ve viyadükler çökebilirken, yollar hasar görebilir veya enkazla kaplanabilir. Bu sık görülen tehlike bazı yol bölümlerinin kapanmasına ve hatta yol ağının bağlantısının kesilmesine neden olur. Afet müdahale aşamasında, çalışma ekipleri, yol ağını yeniden bağlamak ve ulaşıma açmak için yolların bir alt kümesini açmak üzere işe gönderilir. Kapalı bir yoldan, bir takım tarafından bu yol açılmadan geçilemez. Bu çalışmanın amacı, yol temizleme ekipleri için senkronize edilmiş bir çalışma programı üretecek etkin bir çözüm yöntemi sunmaktır. Bloke olmuş yolları temizlemek ve tekrar bağlantı sağlamak için gönderilen birden çok iş ekibinin rotalarını bulan iki ayrıt rotalama problemi ele alınmıştır. Birinci problemde amaç ağın tamamen yeniden bağlanması için gereken zamanı enküçüklemektir. İkinci problemde ise amaç, bağlantısı kesilen ağ bileşenlerini belirli bir süre içinde tekrar bağlayarak kazanılan toplam ödülü enbüyüklemektir. Her bir problem için tam sonuç veren bir karışık tam sayılı programlama (KTP) modeli geliştirilmiştir. İlk problem için KTP modelinin bir gevşetmesinden olurlu çözüm üreten bir mat-sezgisel önerilmiştir. Gevşetme çözümünün optimalite boşluğunun, K'nın takım sayısı olduğu durumda, gevşetilmiş modelden elde edilen alt sınırın en fazla K katı olduğu kanıtlanmıştır. İkinci problem için ise, bir mat-sezgisel yöntem geliştirilmiştir. Bu yöntem, tek araç problemlerini sırasıyla güncellenmiş ödüller ile çözer. Bir üst sınır elde etmek için, önce kesin formülasyondaki zamanlama unsurları gevşetilip, daha sonra Lagrange gevşetme yöntemi ile tek araç problemlerine dönüşen gevşetilmiş KTP çözülür. Önerilen yöntemlerin etkinliği, sistematik şekilde rasgele oluşturulan çeşitli büyüklüklerdeki Öklidyen veriler ve tahmini deprem senaryolarına göre üretilen üç farklı İstanbul yol ağı verisi üzerindeki hesaplamalı deneyler ile gösterilmiştir.
Fiyat dalgalanmalarının bulunduğu ortamlarda envanter yönetimi, fiyatlama ve risk azaltımı
Price uncertainties are among the most critical challenges that retailers and manufacturers have to face. For instance, companies whose operations require procuring from commodity markets are exposed to commodity price fluctuations which experience sharp movements frequently. Besides random nature of customer demand, due to this input and/or sales price volatility, there might be considerable variability in firms' profits. It is vital for these firms to consider price fluctuations in adjusting inventory control and pricing policies, and take a variety of risk management measures. In this dissertation, we consider such a firm where continuous price changes during the planning horizon affect both unit payoff from sales as well as customer arrivals. In a multi-period setting, we first investigate optimal price-dependent inventory control policies and numerically illustrate how continuous price fluctuations affect optimal controls and resulting payoffs. Then, we analyze optimal pricing policies assuming that sales prices are determined both by market-driven random prices and firm's markup decision. We show that level of price variability has a negative effect on firms' final profits. Finally, in a minimum-variance framework, we explore financial hedging strategies of the risk-sensitive firm. We assume that inherent price dynamics of the inventory item are correlated with prices of various products which are freely traded in financial markets. This presents an opportunity for the firm to invest in a financial portfolio of these products to manage its exposure to price and demand uncertainties by observing the current inventory, wealth and price levels. In this environment, we explicitly characterize dynamic variance-minimizing investment decisions of the firm using dynamic programming. We then explore the risk reduction effects of minimum-variance financial hedges through numerical examples and show that significant risk reductions may be possible by using the right hedge.
Navigasyon ve arama konusunda çeşitli çevirimiçi eniyileme problemlerinin analizi
Bu tezde, ağ yapıları üstünde navigasyon ve arama ile ilgili çeşitli çevrimiçi eniyileme problemleri üzerinde çalışılmıştır. Çevrimiçi problemlerde bilgiler adım adım açıklanır ve tüm bilgiler mevcut olmadan önce kararlar alınmalıdır. Tez kapsamında, afete müdahale, arama kurtarma, güvenlik ve savunma alanlarında uygulamaları olan birkaç çevrimiçi eniyileme problemi için eniyi stratejiler tasarlanıp bunların performansları teorik olarak analiz edilmiştir. Önerilen stratejilerin performanslarını analiz etmek için en kötü durumda, bilginin baştan elde olduğu (çevrimdışı) durumdaki en iyi çözüme göre, rekabetçi oranlar belirlenmiştir. İlk olarak, çevrimiçi k-Kanadalı Gezgin Problemi (k-KGP), ayrıtları kesişmeyen yollara sahip çizgeler üzerinde incelenmiştir. Daha önce literatürde bu problem için bir eniyi rassal strateji verilmiştir. Bu çalışmada, bazı durumlarda bu stratejinin uygulanamaz olduğu gösterilerek, strateji her durumda uygulanabilir ve eniyi olacak şekilde değiştirilmiştir. Daha sonra çevrimiçi çok katılımcılı k-KGP ele alınmıştır. Bu problemin sınırlı ve sınırsız iletişimin olduğu iki durumuna bakılarak, literatürde verilen, deterministik stratejilerin rekabetçi oranına alt sınırı iyileştirilmiştir. Aynı iki durum, ayrıtları kesişmeyen yollara sahip çizgeler üzerinde incelenerek, iki deterministik strateji geliştirilmiştir. Bunlardan bir tanesinin eniyi olduğu ispatlanmıştır. Problemin iletişimin olmadığı, sınırlı ve sınırsız iletişimin olduğu üç durumuna bakılarak, rassal stratejilerin rekabetçi oranına alt sınırlar geliştirilmiştir. Ayrıca iletişimli durumlar için rassal bir strateji geliştirilerek, bunun ayrıtları kesişmeyen yollara sahip çizgeler üzerinde eniyi olduğu ispatlanmıştır. Ayrıt belirsizliği olan çevrimiçi Minimum Gecikme Problemi de ele alınan bir başka problemdir. Bu problem için bir eniyi deterministik strateji geliştirilmiştir. Ayrıca, rassal stratejilerin beklenen rekabetçi oranına bir alt sınır bulunmuştur. Son olarak, çevrimiçi Ayrık Arama Problemi, yönlendirilmemiş çizgelerdeki seyahat ve arama maliyetleriyle incelenmiştir. Eniyi deterministik ve rassal stratejiler bulunmuştur.
Çeşitli konveks olmayan problemlerin kopozitif formulasyonlarının dıştan yaklaşımları üzerine
Copositive optimization is linear optimization over the convex cone of copositive or completely positive matrices. "The term copositive programming" was first introduced in 2000. In 2009, Burer showed that mixed binary quadratic optimization problems (MBQP), which comprises a rather large class of nonconvex and combinatorial problems, can be equivalently reformulated as a copositive optimization problem. This seminal work has greatly increased the interest in copositive optimization. It is not surprising, however, that since many combinatorial and nonconvex optimization problems can be reformulated as a copositive optimization problem, copositive programs are also NP-hard in general. The difficulty in the reformulation is entirely due to the conic constraint. For this reason, many researchers have proposed outer approximation hierarchies to the intractable completely positive cone. These approximation hierarchies are composed of a sequence of tractable cones that yield increasingly better approximations of the completely positive cone and are exact in the limit. By replacing the intractable cone by outer approximations in the copositive formulation of nonconvex and NP-hard minimization (resp. maximization) problems, a sequence of increasingly tighter lower (upper) bounds can be obtained for the original problem. This provides opportunities to obtain near-optimal solutions and improve the effectiveness of the algorithms for solving the original problem. In this thesis, we study outer approximations of the copositive reformulations of three classes of nonconvex and NP-hard optimization problems. We first study the class of mixed binary programs (MBPs). We compare the lower bounds arising from outer polyhedral approximations to the lower bound provided by the linear programming (LP) relaxation and establish that the lower bounds due to outer approximations are at least as good as that of LP relaxation. We establish various necessary or sufficient conditions under which the lower bound arising from the outer approximations matches that from the LP relaxation. Our results illustrate the weaknesses of polyhedral approximations. On the other hand, we show that the non-polyhedral doubly nonnegative (DNN) approximations, in general, yield tighter lower bounds. Secondly, we focus on the specific 0-1 knapsack problem (KP) in the class of MBPs. We study two different copositive formulations of the knapsack and compare the upper bounds arising from outer polyhedral approximations to the upper bound provided by the LP relaxation of (KP). We prove that upper bounds obtained from outer polyhedral approximations actually coincide with the upper bound provided by the LP relaxation until at least a certain and fairly large level of the hierarchy. On the other hand, we establish that if the LP relaxation has a non-integer unique solution, then the DNN relaxation gives a strictly better upper bound than the LP relaxation. Finally, we consider the standard quadratic programs (StQP) and investigate the instances of (StQP) for which the DNN relaxation is exact. We establish a complete algebraic characterization of the (StQP) instances that admit an exact DNN relaxation. We explicitly identify three different subsets of such (StQP) instances. Furthermore, we propose a recipe for constructing instances of (StQP) with an exact DNN relaxation. In summary, our results reveal that outer polyhedral approximations, in general, yield weak bounds for (MBP) and for the specific 0-1 knapsack problem, whereas doubly nonnegative relaxations usually give rise to tighter lower bounds.
Üretim sistemlerinin işarete bağlı eşik kuralı ile veriye dayalı kontrolü
With shop-floor data becoming more available, production systems can be controlled more effectively by using data-driven control methods. This thesis focuses on the following research question: how can the decision to produce or not to produce at any time be given depending on the real-time information about a production system?; how can the collected data be used directly in optimizing the policy parameters?; what is the effect of using different information sources on the performance of the system? and how can this choice be made? In order to answer these questions, a production/inventory system that consists of a production stage that produces to stock to meet random demand is considered. The system is not fully observable but partial production and demand information, referred to as markings is available. We propose using the marking-dependent threshold policy to decide whether to produce or not based on the observed markings in addition to the inventory and production status at any given time. An analytical method that uses a matrix geometric approach is developed to analyze a production system controlled with the marking-dependent threshold policy when the production, demand, and information arrivals are modeled as Marked Markovian Arrival Processes. A mixed integer programming formulation is presented to determine the optimal thresholds. Then a mathematical programming formulation that uses the real-time shop-floor data for joint simulation and optimization (JSO) of the system is presented. Using numerical experiments, we compare the performance of the JSO approach to the analytical solutions. The results indicate that the marking-dependent policy introduced here is able to make use of the available partial information effectively and the data-driven joint simulation and optimization is an efficient way for setting the parameters of the policy. Next, we investigate how machine learning methods can be employed to facilitate the optimization of systems controlled with the marking-dependent policies and compare the performance of a number of well-known machine learning methods on this task. Through numerical experiments we show that using machine learning and active learning can improve the performance of complex production systems by speeding up time-consuming optimization problems. Finally, we implement the concepts of this thesis using a Lego production system and discuss the usage of Lego production systems in educational and research related to data-driven control.
Yoğun bakım ünitelerinde tekrar denemeli hasta kabul ve taburcu etme kontrollerin incelenmesi
Intensive Care Units (ICUs) are scarce resources and operate most of the time under high occupancy rates. When faced with limited bed availability, arriving patients are sometimes refused or admitted by discharging an existing patient early, which may result in both increased readmission rates and patient/hospital related costs. Yet, there is not a well-defined admission and discharge control policy to decrease such adverse consequences. In this thesis, we consider a conceptual ICU setting that serves multiple types of patients, which reflects the trade-offs between first-time and recurring patients, between different health stages as well as between the early-discharge and rejection decisions. To do this, we define a so-called readmission orbit whose population consists of patients who are to be readmitted after a previous ICU discharge. There are two major approaches available to represent and analyze such a model: Markov Decision Processes (MDPs) and fluid approximations. Since the conceptual model suffers from the curse of dimensionality, we first consider a simple version of the conceptual model with only one patient type and one health condition, by which we isolate the effects various admission and discharge decisions on readmissions. We present a discrete-time MDP formulation for the simple model and investigate the structure of the optimal admission and discharge control policy under such a model. Next, we present discrete-time MDP formulation for the conceptual model, which aims to develop a well-performing and implementable policy, rather than finding an optimal policy. Then, we develop a deterministic fluid model to control admissions and discharges in such an environment and show that the optimal control of the fluid model in the steady state can be obtained by solving a nonlinear problem. We derive a heuristic policy based on the solution of this nonlinear problem. Finally, we develop a discrete-event simulation platform that represents the ICU system. The platform is quite general, so that it can represent ICUs with different characteristics as well as a variety of control policies. We use this platform to benchmark the performance of the proposed heuristic policies, with those of the two state-of-the-art policies. Numerical results show that our proposed policies significantly outperform the the state-of-the-art policies.
Kanser hastalığında önemli gen kümelerini belirlemek için geliştirilen en iyileme modelleri
Using genomic characterizations of tumours biopsied from cancer patients has a great importance in understanding the formation and progression mechanisms in cancer. Survival analysis is one of the research methods that is used to predict overall survival time of cancer patients and to understand the aforementioned progression mechanisms. High dimensional structure of the genomic characterizations with the limited number of training samples makes survival analysis a challenging task. To be able to identify the survival associated biological mechanisms, cancer-specific pathway/gene set collections can be integrated into machine learning models. Existing approaches usually follow a two-stage approach that either identify predictive genes using a feature selection method and map these selected genes to known pathways/gene sets, or train separate models for each pathway/gene set and try to pick informative ones considering each model's predictive performance. Following such a two-stage approach might result in inefficacy of mapping selected genes to a known biological pathway/gene set due to highly correlated structure between feature groups or including related or very similar pathways/gene sets into the final model due to analyzing each pathway/gene set separately. In this thesis, rather than following such two-stage approaches, we propose machine learning models that can conjointly identify disease related biological mechanisms and perform survival prediction using only these identified biological mechanisms. Our algorithms obtain a sparse set of pathways/gene sets for the survival associated biological mechanisms by eliminating the uninformative ones from the model. We test our algorithms using 20 cancer datasets obtained from The Cancer Genome Atlas and two cancer-specific pathway/gene set collections as input data. We first propose a survival analysis model that integrates pathway/gene set collection into the model using multiple kernel learning. Our algorithm with conjoint modelling approach obtains statistically significantly better or comparable predictive performances against survival random forest (RF) and survival support vector machine (SVM) using significantly fewer gene expression features. Predictive performances of machine learning algorithms can be increased using multitask learning. For this purpose, we extend our multiple kernel learning-based algorithm towards multitask learning. Our multitask learning algorithm both models multiple cancer datasets simultaneously and integrates cancer related biological mechanisms into the machine learning model. The algorithm is able to identify common underlying biological mechanisms for cancer by obtaining better or comparable predictive results against survival RF, survival SVM, and our multiple kernel learning survival analysis algorithm. We also extend our multitask learning algorithm towards task clustering to identify the groups of cancer types that share similar underlying biological mechanisms. To this aim, we propose a unified formulation for task clustering, survival analysis, and knowledge extraction. Our clustering algorithm identifies relevant cancer groups by obtaining statistically significantly better or comparable predictive performances against survival RF, survival SVM, our multiple kernel learning and multitask multiple kernel learning survival analysis algorithms. Numbers of gene expression features and gene sets used by our clustering algorithm are significantly fewer than those of benchmark algorithms. These results show that our methods that identify survival associated biological mechanisms, obtain better or comparable predictive performances against survival analysis methods developed on genomic data without using the pathway/gene set information in the literature. In addition, we prove that survival prediction can be performed using fewer number of gene expression features compared to these benchmark algorithms. We also identify the cancer groups that share similar biological mechanisms without decreasing the predictive power.
Kanser biyolojisi için gen kümesi tabanlı sınıflandırma modelleri
As one of most prevalent and fatal diseases worldwide, cancer has been the focus of biomedical research for many decades. A wide range of different cancers with various causes have been identified and studied, and many treatment methods and drugs have been developed for these cancers. However, many open questions remain in understanding the mechanisms of cancer. With the advent of the high throughput sequencing technologies and the ever-increasing availability of genomic information gathered from cancer patients, researchers have successfully employed genomic information in diagnosis, prognosis and treatment of cancers. Still, given the large number of genes, their considerable correlation and the heterogeneity of cancers, genomic information alone often falls short of generating interpretable models for cancers. To alleviate these effects, pathways can be incorporated into models of cancer. In this thesis, using pathways, we have developed classification methods that produce interpretable models for progression of cancers and survival outcome. Patients usually experience the growth of a number of tumors in their life-time. However, not all these tumors pose a danger to the patient's life. To communicate the severity of the cancer, practitioners and researchers use the staging convention. The stage of a cancer encodes the level of its spread in the neighboring tissue and the body in general, where early-stages are assigned to local tumors and late-stages to the cancers that have spread in the body. Hence, understanding what drives cancers from early- to late-stages is an important question in understanding the mechanisms of cancer. The survival outlook of a patient is a similar measure for the severity of a cancer and understanding the mechanisms that affect the survival chances of a patient is immensely important in devising treatment strategies. In this work, we use multiple kernel learning for integrating pathways into the classification models for cancers. This allows for developing models that are more accurate and far more interpretable compared to models generated by conventional methods that do not use the pathway information. We then extend this method in several directions to improve accuracy and interpretability of the models. In the second part of this thesis, observing that the level of sparsity of the solutions can differ largely from caner to cancer and is often only indirectly affected by the parameters of the methods, we develop a model with an adjustable measure of sparsity and propose efficient solution methods for solving instances of this problem. In the third part of this thesis, given the known similarities between cancers, we develop a framework for building multitask classification models to improve the classification accuracy of cohorts with limited data using the similar cohorts with abundant data. Finally, we employ optimization techniques such as the cutting-plane method and Benders decomposition to improve the algorithmic performance of these methods to the point that they can be applied successfully to large-scale problems stemming from considering several cancers together in a multitask framework. The practical application of these methods is examined by applying them to the two aforementioned classification tasks across the Cancer Genome Atlas datasets for 27 cancer types. The results of these experiments indicate that the incorporation of pathways into these classification problems facilitates generating more interpretable and more accurate models, the similarity of cancers can be leveraged using a multitask framework towards the same goals, and the optimization methods developed here allow for solving these large-scale problems in efficient time using parallelization technologies.
Üçüncü-parti bir lojistik taşıyıcısının karayolu nakliye operasyonlarının optimizasyonu
Planning of daily freight operations is one of the main challenges in transport logistics. Our study originates from a complex real-life routing problem faced by a third-party logistics (3PL) firm. Most of the 3PL firms employ a decentralized approach to ease the planning process and manage the complexity in their operations. This leads to suboptimal decisions due to the lack of integrated use of their resources as well as missed consolidation opportunities. The aim of this study is to develop a solution methodology to optimize the planning of day-to-day road transport operations of the firm in a centralized manner within an acceptable time frame. We address three novel problems related to road freight operations in this thesis. We first examine the less-than-truckload (LTL) planning problem, which involves pickup and delivery orders with deadlines and cross-docking opportunities. We assume that an unlimited number of trucks of different capacities can be hired from the spot market on a short-term basis to fulfill the LTL transportation requests. The objective is to find a minimum cost assignment of orders to feasible route-vehicle pairs while satisfying practical constraints. We propose a decomposition-based matheuristic algorithm to tackle this problem. Our solution approach demonstrates a significant reduction in the freight costs over the long term. The 3PL firm also operates a limited number of dedicated trucks which are either directly owned or hired through a long-term contract. The second problem we address deals with the assignment of dedicated vehicles to less-than-truckload (LTL) and full truckload (FTL) routes. The LTL routes obtained in the first part can be re-scheduled in this phase while the spot-hired vehicles are replaced with dedicated ones to maximize savings via an Mixed Integer Linear Programming (MILP)-based solution approach. We finally examine the last-mile delivery operations of the 3PL firm where different types of products are delivered to customers within a certain geographic region from a central depot. We model this part as a Rich Vehicle Routing Problem (RVRP) with heterogeneous vehicle fleet, time windows, selectiveness and several other practical requirements while allowing multiple trips per vehicle. We propose a parallelized hybrid evolutionary algorithm for solving the problem. The contribution of our study is twofold. On one hand, it creates actual value for the business through reducing freight costs with the proposed solution techniques. On the other hand, it introduces novel problems and solution algorithms to the optimization literature which can be extended to a wide range of industry problems in the future.