Data driven optimization and applications in complex real-life problems

Abstract

Data-driven optimization (DDO) is a methodology that uses data analysis and problem modeling techniques together, whose popularity has increased rapidly in recent years and is expected to continue to increase rapidly in the next 10 years. In DDO, information is obtained from past data by using data analysis and this information is used in modeling the problem and solving the model. Thanks to this method, the assumptions made while modeling the problem are reduced and the model is closer to the problem encountered in real life. In addition, the information obtained allows the model to be restructured to get results in shorter time. The popularity of DDO has increased in recent years and therefore the number of studies in the field of DDO has increased in the literature. However, it is seen that the boundaries of DDO and the conceptual framework of this methodology have not been drawn since this is a new research field. For this reason, in this thesis, the literature in the field of DDO was first examined and a conceptual framework was presented. It is aimed that the proposed conceptual framework will guide studies in the field of DDO and contribute to determining the boundaries of DDO area in the literature. The proposed DDO methodology was applied to four different problems in this thesis study and the results were discussed. The first stage of DDO is obtaining information from data. At this stage, machine learning methods, data mining methods, fuzzy approaches and statistical methods are used. The first problem addressed in this thesis is the problem of predicting picking travel times, which is the most basic model parameter in solving operational warehouse problems. In this thesis study, the time required to pick the items in the orders received in the warehouse was predicted by using the historical picking data of a large automobile spare parts supplier warehouse. In the problem studied, the picking process is carried out by many different pickers and the picking times of each picker vary. Therefore, which picker does the picking directly affects the picking time. For this reason, first of all, pickers were grouped according to their picking characteristics using historical data, and pickers with similar picking characteristics were grouped under the same group. Fuzzy c-Means approach was used for this grouping process. Then, the data set was updated according to the cluster information obtained and picking travel times were predicted using the artificial neural networks method. Predicted picking travel times can be used as model parameters in all operational warehouse problems such as order picking problems, picker assignment problems and picker routing problems. The second problem addressed is the joint order batching, assigning and routing (JOBAR) problem, which is complementary to the first problem studied. This problem is a complex problem in which order batching, picker assigning and picker routing problems, which are three problems encountered in the order picking process in a warehouse, are addressed simultaneously. In fact, all of the lower three problems that make up the problem are NP-Hard in nature, that is, they are problem types where the solution time increases exponentially according to the problem size. Meta-heuristic methods are often used to solve such problems. In this study, the JOBAR problem is addressed in accordance with the DDO methodology. First of all, the picking travel time, which directly affects the outcome of each problem, was predicted using historical picking data with the artificial neural network method. In order to increase the accuracy of this prediction, the pickers' past picking characteristics were taken into consideration and included in the prediction process. The pickers were grouped into four groups according to their past picking performances, and this group information was used as input in the training of artificial neural networks. In the modeling phase, which is the second stage of DDO, the problem was first modeled with the mathematical modeling method and a mathematical model of the JOBAR problem was proposed, covering all three sub-problems. In the proposed mathematical model, picker group information reflecting the picking characteristics of the pickers and piking travel times obtained from artificial neural networks were used as model parameters. However, due to the complex structure of the JOBAR problem, it was not possible to obtain results from the mathematical model within the required time, and therefore a new heuristic algorithm based on the k-nearest neighbor algorithm was proposed for the JOBAR problem. The performance of the proposed heuristic algorithm is compared both with the results obtained from the mathematical model within a certain time limit and with the Clark&Wright algorithm, which is a frequently used savings algorithm in the literature. It has been observed that the proposed new heuristic algorithm is quite successful for JOBAR in terms of both performance and solution time. The third problem addressed with the DDO method in this thesis study is the scheduling crude oil operation (SCOO) problem. As in JOBAR, the SCOO problem is also a complex NP-Hard problem. Although there are different mathematical and meta/heuristic methods proposed for the SCOO problem in the literature, the number of studies dealing with a real-sized SCOO problem is very few. In this study, the SCOO problem was approached and solved with DDO methodology. First of all, by using an event-based mathematical model in the literature, the small sizes SCOO problem was solved for a limited planning period. Then, the result obtained from the mathematical model was analyzed using statistical methods and the Apriori algorithm, one of the most frequently used data mining methods in the literature. A new heuristic method has been proposed for the SCOO problem using the parameters obtained as a result of the data coming from previous stage analysis. The proposed heuristic method has been applied for different planning periods and its performance has been tested. The results obtained show that the proposed approach according to DDO methodology is successful. The last problem addressed in this thesis is the data-driven multi-criteria group decision making (MCGDM) problem. In MCGDM problems, two parameters directly affect the outcome of the decision process. These are the individual effects of the evaluations made by group members on the group decision, that is, the weights of the decision makers and the weights of the evaluation criteria. In this study, a flow is proposed in which the weights of decision makers will be obtained from historical data in accordance with the proposed DDO methodology. In the proposed approach, first of all, the performance of the group members in the past was evaluated and their weights were estimated for the current decision problem. Artificial neural networks were used in this determination process. The proposed methodology was applied on the created synthetic data set and the results were shared. In summary, in this thesis study, a new methodology has been proposed that determines the limits and flow of the DDO approach in the literature. It is aimed to use this proposed methodology in the analysis of the existing literature and to guide future DDO studies. In the continuation of the thesis, the proposed DDO methodology was applied to picking travel time prediction, joint order batching, assigning and routing, crude oil operation planning and multi-criteria group decision making problems, respectively, and the obtained results were shared. The results obtained show that successful results are obtained when the proposed DDO methodology is used as a solution approach for problems.