3 53 MACHINE LEARNING SUPPORTING SUBSTITUTIONS IN SOCCER 1. INTRODUCTION Soccer is a highly competitive and physically demanding sport. The physical demand is highlighted by an increase in ball (game) speed by 15% over the last 50 years [1]. A cohesive body of research points out that players’ fatigue leads to a decline in their running activities. For instance, in a team participating in the Australian national soccer league, total distance, average speed, high-intensity running distance, and very highintensity running distance decreased significantly from the first to the second half by 7.92, 9.47, 10.10, and 10.99%, respectively [2]. In similar fashion, in the Italian A series, a team showed a significant reduction between the first and second half in high-intensity running distance (-14.9%) [3]. These examples highlight that, players are unable to perform maximally throughout a match [4]. Information on this drop in performance is essential for players and coaches. A recent study showed that running performance parameters (e.g., the number of accelerations or decelerations and the distance covered in different speed categories) affect successful performance soccer for some playing positions [5]. As most soccer matches are often decided by just one goal [6], a drop in physical performance can make the difference between winning and losing. Therefore, teams and coaches need to identify players that physically underperform in a match as early as possible to adapt their style of play or substitute these players. In general, an injury of a player, necessary tactical changes, or underperformance of a player causes substitutions (for an overview, see Hills et al., 2018) [7]. Substitution may be the most powerful tool of a coach to influence the match. Substitutions can minimize or offset the effects of fatigue of the team as substitutes cover more distance and perform more high-intensity actions relative to entire match players [8]. According to the Fédération Internationale de Football Association (FIFA) SARS-COVID-19 2020 rules, a coach has five substitution options in a match, implicating fitness of the individual player and physical performance has more impact on substitution than before COVID-19 [9]. To identify a physically underperforming player, coaches can base their decision on realtime motion data. To record and monitor real-time motion data, multi-camera position tracking systems such as SportVU and TRACAB® system are most commonly used in professional leagues [10]. However, one has to constantly monitor and analyse several physical variables of all eleven players. As highlighted in a survey paper by Nosek, Brownlee, Drust, & Andrew (2020) [11], staff and IT solutions struggle with giving helpful feedback to the coach after training sessions due to the amount and complexity of the data and its often inconclusive communication [11]. In order to enable helpful, timely feedback, Robertson advocated using machine learning approaches decision support for the coach [12]. Decision support provides superior efficacy when the volume of the data is large, and the data is complex [13]. An in-match physical performance prediction
RkJQdWJsaXNoZXIy MjY0ODMw