Tıbbi veri sınıflandırması için yapay sinir ağını geliştirmek için meta-heuristik algoritmalar

The tremendous growth of computer hardware technologies and their abilities to solve huge amounts of complex of data has motivated researchers to overcome complicated data mining challenges and problems. Medical dataset classification represents one of the most crucial and complicated problems faced by researches in the field of artificial intelligence and data mining. The different diseases and the various ways of diagnosis by using multiple testing have produced large amounts of complex medical data. Moreover, the huge number of patient records in clinical centers and hospitals and other health institutions has generated the need for advanced and accurate medical mining applications to help doctors and therapists investigate cases regardless whether patients are in critical conditions or require remote follow-ups. This thesis focuses on the hybridization of the artificial neural network (ANN) and metaheuristic algorithms to enhance the accuracy of a classification model for the overlapping fields of medical data mining. The key problems associated with medical diagnoses involve the identification of highly accurate classification models. The contributions of this thesis revolve around the two important classification problems or issues highlighted in the related literature. For the first strategy, the relation between the ANN structure and the optimized algorithm is established. For the second strategy, the tradeoff between diversification and intensification is investigated as part of the search for the optimal global solution. In the first chapter we discuss a background introduction and in the second chapter a literature survey about the approaches applied on the problem. The third chapter of this thesis discusses the effect of metaheuristic iteration on ANN structure. The novelty of the proposed work shown through improving the impact of metaheuristic algorithm iteration on ANN structure. ANN is enhanced using separate three metaheuristic algorithms (particle swarm optimization PSO, genetic algorithm GA, and fireworks algorithm FW). The proposed models are tested on five standard medical benchmarks and one big-data medical dataset. The proposed study is successfully implemented, and remarkable results are obtained. Furthermore, the no-free-lunch theorem (NFLT) is verified in the study's context, that is, no algorithm is universal for all problem domains. The forth chapter of this thesis investigates the tradeoff between exploration and exploitation when obtaining the optimal global solution which represent best accuracy of medical data classification. The number of hidden layers and the number of neurons in each layer can both affect ANN learning. Thus, the ANN used in this thesis involves the selection of a complex structure that can achieve highly accurate results; consequently, metaheuristic algorithm efficiency can be guaranteed when searching for the global optimum. Two metaheuristic algorithms named differential evolution algorithm DE and simulated annealing SA are combined to formulate a new and improved algorithm DESA for considered problem domain. However, selecting the highly accurate two algorithms is not mandatory; instead, empirical tests can be performed for convenience. Originality of proposed method is combining between DE as evolutionary metaheuristic algorithm and SA as trajectory algorithm to provide balance between exploration and exploitation to explore search space widely for global solutions and intensively exploited local solutions. DESA method compared with tow evolutionary which are GA and DE, and with two trajectory which are SA and Tabu search TS. Proposed method DESA is implemented successfully, and better results obtained.