Processes in sport can be described as time series of patterns, which can as well characterize situations (e.g. positions on the playground or angles of articulations) as activities (e.g. moving of players or angle speeds). Patterns can be learned and recognized by means of "self organizing maps" (SOM) the most famous type of which is that of Kohonen Feature Map (KFM) (see Kohonen (1981), Hopfield (1982), Köhle (1990), Polani & Uthmann (1993)). Therefore SOMs like KFMs can help to analyse processes in sport, as has been done in several approaches (see Lames & Perl (1999), Schöllhorn & Perl (2002), Schöllhorn et al. (2002), Lippolt et al. (2004)). However, there is a type of problem that is difficult to handle with a "conventional" KFM - namely if learning itself is the process to be analysed: Due to the fact that a KFM learning process is controlled by an external algorithm using parameters that run down to final values and so eventually cause the end of the learning process, a once trained KFM cannot be reactivated. Therefore additional or continuing learning can be done only by repetitions of the learning process using appropriate mixtures of data from the different phases of the learning process - which is uncomfortable as well as methodologically not satisfying. In order to handle this problem, the concept of Dynamically Controlled Network (DyCoN) has been developed in our working group.
© Copyright 2004 International Journal of Computer Science in Sport. Sciendo. Kaikki oikeudet pidätetään. / Kääntänyt https://www.DeepL.com/Translator