Ümmügülsüm Çelik

Artificial neural network (ANN) is one of the machine learning algorithms widely used in prediction studies recently. The key to obtaining effective prediction results with ANN depends on its training and the design of its tunable parameters. The Backpropagation and Levenberg-Marquardt (BP-LM) learning algorithms are the most utilized algorithms for training ANN. However, these algorithms have some disadvantages such as local minima, computational complexity, sensitivity to initialization, overfitting, and limited parallelism. In this study, we proposed a Particle Swarm Optimization (PSO)-trained ANN model to deal with these problems in ANN learning. PSO is one of the well-utilized artificial intelligence algorithms and it can be successful in the learning process thanks to its features of finding global optima, having a few parameters to be tuned, being easily parallelized, robustness and convergence speed. The proposed model is tested with different ANN structures and parameter values for tourism revenue prediction. As a result, it was observed that the proposed PSO-trained ANN model generally gave better prediction results than BP-LM trained ANN and an optimal ANN structure was obtained for the prediction of tourism revenues. In addition, forecasting of tourism revenues for the next 12 months was obtained with a designed optimal ANN structure.