Updated on 09 Jul. 2008

 

 

 

Keynote abstracts

Analysing learning approaches by means of complex movement pattern analysisDaniel Janssen, Hendrik Beckmann, Florian Gebkenjans, & Wolfgang I. Schöllhorn
University of Mainz; University of Münster / Germany

  

Does the order of movement executions have effects on the movement variability? We investigated these effects in executing ten first straight and ten second topspin serves in tennis, following the concepts of learning with low and high contextual interference (Brady, 2004). A kinematic analysis was conducted using Selforganizing Maps (Kohonen, 2001). The aim of the study was to answer three questions of interest: i) Is a Selforganizing map (SOM) able to distinguish individuals by their kinematic serve patterns? ii) Is the SOM able to distinguish between an individual’s first and second serve on a finer level respectively? And iii) What effects does the order of movement execution have on the movement variability and precision?

Methods
Five tennis players (regional level) aged 26 ± 1,4 years participated voluntarily in the study. On two different days, participants had to perform 20 tennis serves in each case (10 first serves and 10 second serves). On one day, participants performed first and second serves under low contextual interference (blocked condition: 10 first serves, afterwards 10 second serves), and serves under high contextual interference (serial condition: first and second serve alternating) another day. The participants’ task was to place the serves with maximum speed and precision as near as possible to a target in the diagonal opposite. During performance participants were filmed with two orthogonal positioned high speed cameras with a frequency of 200 Hertz. With the aid of the software SIMI Motion (SIMI Motion, Germany), joint angles of a 14-segment body model were calculated and exported to Matlab R2007b (The Mathworks, USA). For further analysis the SOM Toolbox (Vesanto, Himberg, Alhoniemi, & Parhankangas (2000) was used. As the input dimension of the kinematic data (15 angles) was considerably high, for questions i) and ii) a self-implemented network called 2-SOM (due to its structure that consisted of two series-connected SOMs) was chosen for analysis purposes. Within the 2-SOM, the first SOM (SOM A) had the task of reducing the data dimension, whereas the second SOM (SOM B) took care of the classification of serve patterns. This procedure is based on the original work of Bauer and Schöllhorn (1997) and Barton, Lees, Lisboa, & Attfield (2006). An additional possibility of testing recognition rates was implemented within the 2-SOM. Therefore cross-validation was used (Schöllhorn, Jäger, & Janssen, 2008), testing the trained network in each step with unknown data from the dataset. This approach is estimated to deliver worse recognition rates than multilayer perceptrons for example, but due to its hypothesis generating character, it is favoured in the current study. For question iii) a custom algorithm was designed that compared mathematical differences between the serve patterns in a manner similar to Euclidian distance calculation.

Results
A custom 2-SOM revealed a person recognition rate of 100% using all available data from all participants. In consequence, concerning the data for only the blocked or serial condition, the same network delivered 100% recognition rates as well. However, within the person clusters, that were formed in the output space of these maps, a network that was trained whether with all individuals’ data from the blocked or serial condition, was able to distinguish between an individual’s first and second serve with an accuracy of 73.0% (blocked condition) and 66.67% (serial condition) respectively.

   Intra-individual recognition rates of first and second serve were calculated as well and were about 95.56% (blocked condition) and 94.44% (serial condition) on average. Therefore a new smaller network was designed, that classified the data of the participants separately.
For the second serve a nearly equal distributed variance in the upper and lower parts of the participants’ body can be found in both training conditions, whereas for the first serve these differences are of greater magnitude. Furthermore, in the blocked condition notably more variance can be observed in the lower body parts, whereas in the serial training condition more variance is existent in the upper parts of the body.
Concerning the movement variability, no variability differences between the second serve in the blocked and serial condition were found. However, movement variability in the first serve was considerably increased for the serial condition in comparison with the blocked condition, although these findings could not be confirmed statistically due to the small group size (see Fig. 1).
Fig. 1: Movement variability in the first and second serve under low and high contextual interference, averaged over all participants.

Target precision was analysed as well. The mean standard deviation from the target was 0.61 meters for the blocked and 0.68 meters for the serial condition. The single results in the blocked condition were 0.64 m (first serve) and 0.53 m (second serve). In the serial condition the difference between the two serves was smaller (0.66 m for the first, and 0.64 m for the second serve).

Discussion
Self-Organizing maps have shown their ability to recognize individual movement patterns with recognition rates of up to 100%. On an intra-individual level, SOMs succeeded in distinguishing first from second serves with up to 95.56%. Movement variability was higher for the first serve in serial condition in comparison to the blocked condition, whereas no differences could be found for the second serve in both conditions. The best target precision was achieved with the second serve in the blocked condition. The differences in target precision were much smaller for the serial condition. According to this results the contextual interference approach can considered as a subset of the differential training approach.

Literature
Barton, G., Lees, A., Lisboa, P., & Attfield, S. (2006). Visualisation of gait data with Kohonen self-organising neural maps. Gait & Posture, 24(1), 46-53.
Bauer, H., & Schöllhorn, W. (1997). Self-organizing maps for the analysis of complex movement patterns. Neural Processing Letters, 5(3), 193-199.
Brady, F. (2004). Contextual interference: a meta-analytic study. Percept Mot Skills, 99(1), 116-26.
Kohonen, T. (2001). Self-Organizing Maps. Berlin: Springer.
Schöllhorn, W. I., Jäger, J. M., & Janssen, D. (2008). Artificial neural network models of sports motions. In Y.B. Hong, R. Bartlett (Ed.), Routledge Handbook of Biomechanics and Human Movement Science. (pp. 50-64). Routledge: London.
Vesanto, J., Himberg, J., Alhoniemi, E., & Parhankangas, J. (2000). SOM Toolbox for Matlab 5. Report A57. Helsinki University of Technology. Neural Networks Research Centre. Espoo.

© copyright IDRAM - 2008