SIMULATION OF THE FLIGHT DISTANCES OF JAVELINS BASED ON A NEURAL NETWORK APPROACH

  • Klaus D. Maier
  • Veit Wank
  • K. Bartonietz
Keywords: javelin throwing, neural networks, flight simulation

Abstract

INTRODUCTION: The flight distances of javelins are determined by the release parameters as well as by the forces acting on the javelin during flight. The flight phase of the javelin has been under investigation by many researchers using engineering approaches to model the flight phase. The objective is to allow an optimization of the release parameters for maximizing the flight distance. The measurement of release parameters as well as wind influence is not very precise. This means that the models are based on already distorted data. Artificial neural networks (NNs, Haykin 1994) are powerful information processing tools that allow to construct a input-output model of a problem by learning from examples. They are able to generalize , i.e. to produce reasonable outputs for inputs that have not been encountered during learning. NNs handle imprecise data well and could be suitable for modeling the flight distance of javelins as a result of the release parameters. METHODS: Release parameters have been measured using three dimensional film and video analysis. Relevant parameters were determined: the angle of release, the angle of attack (seen from the side), the angle of side attack (seen from behind) as well as the velocity of release. The overall flight was measured as the distance between the throwing line and the athlete’s hand at the point of release plus the distance between the line and the point of touch down of the javelin. Other parameters such as javelin brand, wind speed, etc., were not considered in the model. Multi-Layer-Perceptron Neural Networks (MLPs) were used to construct a model with the release parameters as inputs and the overall distance as output. RESULTS: Several setups were used for the training of the MLPs and 40 sets of release parameters were processed. We used 37 sets for the training of the MLPs and 3 sets were kept for examining the MLPs’ generalization performance (crossvalidation). This was repeated with randomly selected sets for training and crossvalidation. Predictions of the total flight distance using the release parameters were exact up to 5 percent of the overall distance for the cross validation sets. CONCLUSIONS: The MLP simulation of the flight distance is a suitable instrument even though it uses only a small number of parameters. This can be helpful for coaching and provides an alternative to other models. Using more data sets may improve the quality of prediction, and further work will include recording more data sets as well as studies on optimal javelin release parameters. REFERENCES: Haykin, S. (1994). Neural Networks. Englewood Cliffs: Macmillan Publishing Company.