Comparison of real judo competition results with artificial neural networks method

In this study, it is aimed to predict Judo competition results with artificial neural networks. For this purpose, 21 different models were created with hyperparameters of the number of layers, number of neurons and optimization methods (SGD, RMSprop, Adam, Adadelta, Adagrad, Adamax, Nadam) in the artificial neural network. The sample of the study consists of 7758 athletes competing in international judo competitions between 2017:01-2021:03. 14 different attributes of each athlete were calculated from the data obtained from 53775 judo competitions held in this period. By sending the attributes of two contestants to the input layer of the neural network, 28 input data and 1 output data were created. The application is trained with seven different optimization methods in the neural network with 64 neurons in one hidden layer, 32-64 neurons in two hidden layers, and 64-128-64 neurons in three hidden layers, respectively. As a result of the application, it was determined that the most successful model (78.6% accuracy, 44.4% error) used 64 neurons in a single layer, RMSprop optimization method. It was determined that the model with the lowest success rate (74.1% accuracy, 51.8% error) used the Adadelta optimization method with 32-64 neurons in its two hidden layers. It was determined that the optimization methods RMSprop and Adamax were more successful than the other methods, while the Adadelta method was more unsuccessful. As a result, it has been revealed that Judo competition results can be predicted with artificial neural networks by using appropriate dataset and hyperparameters.