Training of artificial neural networks via modern meta-huristic algorithms

The purpose of this thesis is to provide ANN training, which is a good classification tool used in optimization problems. In accordance with the purpose of the thesis, Multilayer Perceptron (MLP) is used from ANN architectures. Modern meta-heuristic algorithms which are probabilistic algorithm types in MLP training are selected as trainers. Ackley, Easom, Egg Crate, Griewank, Quartic, Rosenbrock, Schwefel, Sphere and Step bencmark functions are used to measure the overall performance of the algorithms and to compare performance advantage. XOR, Iris, Breast Cancer, Heart Diseases and Balloon datasets were taken from the UCI database to demonstrate the closeness of the classification rates to reality in MLP education. Four applications were made in the thesis. In the first application, the advantages of Harris Hawks Optimization algorithm (HHO) were compared with other algorithms. In the second application, the performance advantages of the Atom Search Algorithm (ASO) has been statistically demonstrated by showing the solution of the global constrained optimization problems of the modern meta-heuristic algorithms. In the third application, the Sine Cosine Algorithm (SCA) and HHO algorithms were compared by making MLP training. In the fourth aplication, through proposed hybrid ASO-SA algorithm (hASO-SA) has been to eliminated by the disadvantages of the meta-heuristic algorithms in the problem solution. By showing the performance advantage statistically, high classification skill in MLP education is explained by comparing with other meta-heuristic algorithms.