An Example of Classification Using a Neural Network Trained by the Zebra Optimization Algorithm

Artificial intelligence techniques are a broad field of research with training, computation and prediction capabilities. Among these techniques, artificial neural networks (ANNs) are widely used as a predictive model. Learning algorithms in ANN classifiers have great importance on the success of ANN. The ANN model generally uses gradient-based learning models. However, due to the disadvantages of gradient-based learning models in local search, they have begun to be replaced by heuristic-based algorithms in recent years. Heuristic algorithms have attracted the attention of many researchers in recent years due to their success in problem solving. In this study, the Zebra Optimization Algorithm (ZOA), which has been proposed recently to train ANN networks, was examined. The main purpose of this study is to train the neural network using ZOA and increase the sensitivity of the perceptron neural network. In this study, a new ANN network integrated with ZOA is proposed. In this study, a detailed parameter analysis was carried out to show the effect of the population size and maximum generation number parameter settings, which form the basis for ZOA, on the ANN network. Then, a parameter analysis was carried out for the number of layers, number of neurons and epoch values, which are important for ANN networks. Such an ideal ANN network has been identified. This ideal ANN model was run on seven different data sets and was successful in predicting accurate data. In addition, three different heuristic algorithms (Gazelle Optimization Algorithm (GOA), Prairie Dogs Optimization (PDO), and Osprey Optimization Algorithm (OOA)) selected from the literature were integrated on the same ANN model and compared with the results of ANN integrated with ZOA operated under similar conditions. The results reveal that the proposed algorithm leads to greater convergence with the neural network coefficient compared to other algorithms. In addition, the proposed method caused the prediction error in the neural network to decrease.