Abstract des Autors

The paper presents the use of linear and nonlinear multivariable models as tools to predict the results of 400-metres hurdles races in two different time frames. The constructed models predict the results obtained by a competitor with suggested training loads for a selected training phase or for an annual training cycle. All the models were constructed using the training data of 21 athletes from the Polish National Team. The athletes were characterized by a high level of performance (score for 400 metre hurdles: 51.26+/-1.24 s). The linear methods of analysis include: classical model of ordinary least squares (OLS) regression and regularized methods such as ridge regression, LASSO regression. The nonlinear methods include: artificial neural networks as multilayer perceptron (MLP) and radial basis function (RBF) network. In order to compare and choose the best model leave-one-out cross-validation (LOOCV) is used. The outcome of the studies shows that Lasso shrinkage regression is the best linear model for predicting the results in both analysed time frames. The prediction error for a training period was at the level of 0.69 s, whereas for the annual training cycle was at the level of 0.39 s. Application of artificial neural network methods failed to correct the prediction error. The best neural network predicted the result with an error of 0.72 s for training periods and 0.74 for annual training cycle. Additionally, for both training frames the optimal set of predictors was calculated.