The prediction of action positions in team handball by non-linear hybrid neural networks

One key to success in sport games is to anticipate individual and team tactical behaviour. The earlier and the more accurate players can identify the opponents’ strategy, the higher is the probability of success. To date it is not known if artificial intelligence is able to perform such anticipations in sports games. Therefore, the aim of the present study is to predict the shooting in team handball by means of artificial neural networks (ANN) in two approaches. In approach 1 only the shooting position was predicted and in approach 2 the last pass preceding the shot and the shot position itself was predicted from position data from preceding actions. Position data from 723 actions sequences of the EHF U18 Men European handball championship were annotated. A hybrid prediction system radial basis function network was used for prediction. Seventy per cent of the data-set were used for the network training, 15% for cross-validation and 15% for testing. The ANN predicted the real positions with an accuracy of 1.20 ± 0.46 and 1.42 ± 0.77 m for approach 1 and 2, respectively. Results demonstrate that ANNs are capable to predict position data in the sports game team handball with meaningful accuracy.