Artificial neural networks predicting the outcome of a throwing task : effects of input quantity and quality
Deutscher übersetzter Titel: | Künstliche neuronale Netze zur Prognose des Ergebnisses bei einer Wurfaufgabe : Effekte von Inputquantität und -qualität |
---|---|
Autor: | Joch, Michael; Jäger, Jörg M.; Maurer, Heiko; Maurer, Lisa Katharina; Müller, Hermann |
Erschienen in: | Proceedings of the 11th International Symposium on Computer Science in Sport (IACSS 2017) |
Veröffentlicht: | Cham: Springer International Publishing (Verlag), 2018, S. 23-34, Lit. |
Beteiligte Körperschaft: | International Symposium of Computer Science in Sport |
Format: | Literatur (SPOLIT) |
Publikationstyp: | Sammelwerksbeitrag |
Medienart: | Elektronische Ressource (online) Gedruckte Ressource |
Dokumententyp: | Tagungsband |
Sprache: | Englisch |
DOI: | 10.1007/978-3-319-67846-7_3 |
Schlagworte: | |
Online Zugang: | |
Erfassungsnummer: | PU201807004656 |
Quelle: | BISp |
Abstract des Autors
Internal forward models are used to explain motor prediction processes in motor control and learning e.g. predicting an upcoming miss in a throwing task before the knowledge of results is available. In this study we used artificial neural networks (ANN) to model such movement outcome prediction processes. Additionally, we varied the inputs of four different multilayer perceptrons (MLP) with respect to the quantity and the reliability (quality) of input parameters to account for perceptual noise. The results show that ANNs are able to learn the non-linear input-output mapping underlying the throwing task even with few input variables (velocity and angle at ball release). Results improve when providing additional information about the ball flight (prediction accuracy increases from RMSE = 7.9 mm to RMSE = 3.9 mm). However, when a model is provided with noisy inputs only, model training and prediction suffers substantially (RMSE = 53.8 mm). Yet, additional reliable information about the ball flight (in addition to noisy velocity and angle) leads to very high model prediction accuracy again (RMSE = 4.1 mm). In a nutshell, ANNs can be used to model internal forward model predictions, but the availability of reliable input information is essential at least to some extent.