Introduction: The tactical analysis of player behaviour in team sports like association football is of great interest. The increasing availability of position data allows for an in-depth analysis of tactical behaviours of individual players (Leser et al., 2015). The high complexity of player interactions, however, makes the assessment of individual tactic skills based on time-continues position data difficult. Methods: We address the problem of designing an explanatory computational model for player assessment by fusing fuzzy human-like knowledge related to tactical behaviour with unsupervised feature extraction based on time-continuous position data from a tracking system. Our hierarchical model architecture consists of two layers. The first layer provides an abstract representation of the time domain via temporal segmentation of the data into game-situation specific temporal phases. Conceptually the coarsened view on the time domain is modelled via a Markov chain where transitions between states represent meaningful events. In a second step we then identify different behaviour patterns hidden in the data with unsupervised machine learning technics. The main focus of this layer is to provide the trainer with a qualitative description of the key differences of the behaviour patterns based on key performance indicators. We show how to reduce the high dimensionality of the input space using unsupervised dimensionality reduction methods and how to provide a meaningful explanatory model for the trainer using natural language expressions. Results and Discussion: We illustrate our approach by means of three different kinds of small sided games in association football with increasing complexity. Starting with a 1vs1 situation with low complexity, we show the principals of our modelling approach. Next we show that for 3vs2 and 5vs5 situations our approach is able to identify the underlying dynamical patterns of player interactions. With the multifactorial statistical analysis, as presented in this paper, we are able now to present the complex interaction between the automatically selected measurement variables in a qualitative way that is suitable for the trainer.
© Copyright 2016 21st Annual Congress of the European College of Sport Science (ECSS), Vienna, 6. -9. July 2016. Veröffentlicht von University of Vienna. Alle Rechte vorbehalten.