Human Movement Science 31 (2012) 334–343

Contents lists available at ScienceDirect

Human Movement Science

journal homepage: www.elsevier.com/locate/humov

Tactical pattern recognition in soccer games by means

of special self-organizing maps q
Andreas Grunz a,⇑, Daniel Memmert a, Jürgen Perl b
a
German Sports University Cologne, Institute of Cognitive and Team/Racket Sport Research, Germany
b
University Mainz, Institute of Informatics, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Increasing amounts of data are collected in sports due to techno-
Available online 12 June 2011 logical progress. From a typical soccer game, for instance, the posi-
tions of the 22 players and the ball can be recorded 25 times per
PsycINFO classification: second, resulting in approximately 135.000 datasets. Without com-
4160
putational assistance it is almost impossible to extract relevant
Keywords: information from the complete data. This contribution introduces
Pattern recognition a hierarchical architecture of artificial neural networks to find
Team sport tactical patterns in those positional data. The results from the
Neural networks classification using the hierarchical setup were compared to the
results gained by an expert manually classifying the different cat-
egories. Short and long game initiations can be detected with rela-
tive high accuracy leading to the conclusion that the hierarchical
architecture is capable of recognizing different tactical patterns
and variations in these patterns. Remaining problems are discussed
and ideas concerning further improvements of classification are
indicated.
Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction

Different computer-based approaches try to extract and analyze tactical patterns in sport games.
Besides traditional statistical tools (cf. Hughes & Franks, 2009; Reilly, Cabri, & Araujo, 2005), neural
network approaches from the area of computer science are discussed and evaluated in the field of

q
This article has been supervised by Peter Beek (Editor-in-Chief).
⇑ Corresponding author. Address: Am Sportpark Müngersdorf 6, IG I, 3. Etage, 50933 Köln, Germany. Tel.: +49 (0) 221 49 82
4284.
E-mail addresses: grunz@uni-mainz.de, a.grunz@dshs-koeln.de (A. Grunz).

0167-9457/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.humov.2011.02.008
 A. Grunz et al. / Human Movement Science 31 (2012) 334–343 335

team sport (Grunz, Memmert, & Perl, 2009; Jäger, Perl, & Schöllhorn, 2007; Memmert & Perl, 2006;
Perl, Memmert, Bischof, & Gerharz, 2006; Pfeiffer & Perl, 2006). Considering the problem of recogniz-
ing different variations in tactical pattern several difficulties arise for all approaches.
This paper is concerned with a classification of short and long game initiations which are roughly
defined at first. Both tactics start by winning the ball from the opposing team by the goalkeeper or a
defense player. Next, the ball is passed to different team players. If each pass length exceeds approx-
imately 30 m the complete tactical pattern is classified as a long game initiation, otherwise it is clas-
sified as a short game opening. The tactical pattern ends with losing the ball to the opposite team. It is
a difficult task to formulate precise definitions. Involved players as well as the actions they choose
may vary, leading to an enormous variety of observable patterns. Approaching the issue with a set
of different rules to map all such variations to a representative type seems infeasible. Neither is it pos-
sible to formulate all required rules, nor to guarantee consistency between different sets of rules for
different tactical patterns.
In order to improve classification we use artificial neural networks as a kind of machine learning
approach, replacing the construction of such explicit rules by fuzzy similarity relations, which shows
satisfying results in comparable approaches (Perl, 2002, 2004, 2008). In the following, we will first give
a short introduction to the method of self-organizing maps (SOM), completed by some extensions of
the standard SOM approach. Here, we will give a brief overview of the standard Kohonen’s SOM. Then
we will show the hierarchical architecture in more depth and take a closer look at the different levels.
Therefore, we compare the results from our classification using the hierarchical setup to the results
gained by an expert manually classifying the different categories.

1.1. The nature of SOM: Theoretical background and technical framework

A SOM consists of a set of artificial neurons which are normally arranged in a rectangular grid form-
ing a two-dimensional map (for a review, see Kohonen (1995)). At the beginning, each neuron is ini-
tialized with a random vector whose dimension is equal to the dimension of the multidimensional
input space. This vector is called the neuron’s weight vector. To simplify the notation a neuron
together with its corresponding weight vector will be referred to as a prototype. Before a network
can be used for typification, it must learn the mapping. For this reason a training phase has to precede
the testing or production phase. During training the net is fed with a set of samples, which are drawn
at random from the input space. For each entered data sample the nearest prototypes are slightly
moved towards the direction of the data. How far a prototype is moved depends on the similarity
to the current data sample and the training step – i.e., when training starts the prototypes can be
moved over larger distances while in the final phase a prototype should only make small movements.
Similarity can be measured in different ways; we use the Euclidean distance to measure similarity.
After training the map approximates a two-dimensional nonlinear manifold embedded in an usually
high dimensional data space (Bishop, 2007).

1.2. Dynamically Controlled Network (DyCoN)

The standard Kohonen SOM suffers from one limitation which is relevant for our research. After the
net has stabilized, it cannot be changed anymore. In particular the net cannot learn any new pattern or
new variations in a pattern. Different approaches have been developed to solve this limitation (Perl,
2002, 2004). Dynamically Controlled Network (DyCoN; Perl, 2004) extends the standard SOM with
an adaptive learning rate and distance. Using a Performance Potential (PerPot) Metamodel each neu-
ron controls its own learning behavior enabling continuous learning (for a review, see Perl (2002); for
practical implications, see Memmert and Perl (2009a)).
A SOM and derivations like DyCoN have the very important property that as far as possible similar
inputs, e.g., variations of a pattern are mapped to the same neuron or to neighboring neurons (local
conservation of topology). This is a rather abstract description and it can be better explained with
an example.
336 A. Grunz et al. / Human Movement Science 31 (2012) 334–343

1.3. Example: Training of defense player constellations

Positional data, which have been drawn from a soccer game, form our database for further inves-
tigations. Each position on the field is represented by x- and y-coordinates. First, we group the coor-
dinates of the four defense players in an eight dimensional vector (Fig. 1).
In the following we will refer to such a vector as a constellation. These constellations are used to
train a DyCoN. After training, each neuron encodes a different constellation of four positions on the
soccer field. Obviously, the original weight vector does not tell us much about the constellation it cor-
responds to. Therefore, in a somewhat unusual way, Fig. 2 gives a presentation, where each neuron
makes the encoded constellation directly visible. In our implementation we use, of course, larger
networks, which normally consist of 20  20 neurons.
Note that the encoded constellations do not necessarily represent real constellations but can be
seen as prototypes for groups of input constellations. In this way all possible constellations of the four
players on the soccer field are mapped to a small number of prototypes. Therefore the process of map-
ping real constellations to prototypes is a kind of fuzzification, which makes this approach less vulner-
able to imprecise or erroneous positional data.
As stated, the mapping from data to neurons tries to preserve similarities that exist in the input
data, i.e., similar constellations should be encoded to the same or neighboring neurons. In the current
context, however, similarity has two aspects: on the one hand, two constellations can be found on
neighboring positions on the field (Fig. 3); on the other hand, two constellations can have related
shapes.
Due to dimensionality discrepancy between the input data, which usually has a dimension greater
than 2, and the two-dimensional network not all similarities can be equally preserved. Also, attention
should be paid to the intrinsic dimension of data that can be smaller than the vector dimension. For
instance, if 3 of 4 defense players never move, the intrinsic dimension of data would be 2 instead of 8.
In this unrealistic case we get a 1-to-1 mapping with perfect topology preservation. With increasing
difference between the dimensions more and more similarities in the data are lost. In Fig. 2 both ef-
fects can be observed. In the first row the neurons encode similar constellations and they also are
neighbors. In the second row there is a topology break between neurons 3 and 4. Although they are
neighbors they represent quite different constellations. A third case of topology break, which is not
evident in Fig. 2, is also possible: neurons encoding similar constellations can be positioned far away
from each other on the grid. This is due to a kind of toroidal topology that, e.g., can connect neurons
from different borders of the neurons grid by similarity relations.
A more technical aspect of optimally organizing neurons is local zooming: the mapping from data
to neurons is automatically adapted to the resolution of the data. Concerning our example from soccer

Fig. 1. The positions of the four (Italian) defense players are grouped into a constellation.
 A. Grunz et al. / Human Movement Science 31 (2012) 334–343 337

Fig. 2. For each neuron the encoded constellation is presented in the middle of the corresponding quad.

this approach has two advantages. On one hand the region the defense players mainly act in is repre-
sented by more neurons than the region the players only rarely visit. On the other hand, frequent
formations are represented by more neurons and therefore with a higher resolution than rare ones.
The aforementioned observation of local similarity preservation causes a structuring of the neu-
rons: Neurons encoding similar prototypes can be grouped together to form a cluster. Thereby each
cluster represents a set of similar input data. Moreover, we get a disjoint partitioning of the neural net-
work landscape in regions which represent different types of input data. Fig. 2 shows the different
clusters. Neurons belonging to a cluster of equal grayscale show the same number below the field.

1.4. Example: Test of defense player constellations

Clustering finishes the training phase and prepares the net for the testing phase, which starts with
pre-processing the original input data by forming the needed input constellations from the positional
338 A. Grunz et al. / Human Movement Science 31 (2012) 334–343

Fig. 3. Both constellation are equally shaped but are positioned on different places on the soccer field.

data. After this step the sequence of constellations can be entered into the trained network. For each
data entry following the time course the neuron encoding the most similar prototype is activated.
Thereby the sequence of input data is transformed to a sequence of activated neurons. We can graph-
ically illustrate the transformed sequence by connecting activated neurons with a line (see Fig. 4). In
the following, we will refer to this line as a trajectory.
Trajectories represent the movements of the constellations on the playground and therefore map
the process of the game. Obviously it makes sense to take those trajectories for a net-based analysis
in order to identify types of game processes. Unfortunately, as mentioned above, networks need input
vectors of equal length for training and testing. Moreover, such vectors have to contain just numbers.
Using a sliding window technique we can extract sequences of fixed length from the trajectory. Each
such sequence can be transformed into a vector of numbers by simply taking the numbers of those
clusters which the trajectory neurons belong to (see Figs. 5 and 6).

1.5. Hierarchical architecture of Dynamically Controlled Networks

These trajectory vectors can be used to train and test a second DyCoN, which eventually leads to a
hierarchical network structure that contains networks for offense and defense situations on the con-
stellation level and networks for the moving processes on the trajectory level.
Of course, the information flow through such a complex structure is not quite transparent. There-
fore, after implementation of the hierarchical architecture, first preliminary classification results1
showed poor classification accuracy, which was mainly caused by specific technical problems. As an
example, a deeper analysis revealed that too much information was lost on the typification2 of constel-
lations due to dimensionality mismatch. We therefore divided the x-y-constellations into x-coordinates
and y-coordinates and defined separate vectors for x and y, respectively. In the following, these vectors
will be referred to as x-constellations and y-constellations. In Fig. 7 the complete hierarchical architec-
ture for classification is presented. The arrows indicate the direction of dataflow. On the first horizontal
layer we enter the different x- and y-constellations. On the vertical three different blocks can be distin-
guished: On the left hand side we enter the x- and y-constellations of, e.g., the four defense players; in
the center we supply the interaction counterpart. Here we enter the x- and y-constellations formed of the
position of the four opposite attack players.

1
In this contribution the term ‘classification’ is used in the context of the general aim to find different tactical patterns in soccer
games.
2
The term ‘typification’ – in contrast to the term ‘classification’ – is used in the smaller context of an artificial neural network
characterizing a mapping from an input sample to an artificial neuron.
 A. Grunz et al. / Human Movement Science 31 (2012) 334–343 339

Fig. 4. The sequence of activated neurons is graphically illustrated by a line running over the network (trajectory).

The first layer is responsible for typifying the different constellations. The ball is specially treated in
that there is no need to typify its position, which only consists of one x- and one y-coordinate. After
training we can proceed with the trajectories by constructing appropriate output vectors for the next
layer. On the second layer a pair of two networks describes the movement of the corresponding typ-
ified constellations. The first network characterizes the relative movements of the four x-coordinates.
The relative movements of the four y-coordinates are treated by the second network. Altogether the
movement of the four defense players is described by the two networks in layer two on the left hand
side of Fig. 7.
The detection of tactical patterns requires a complete typified description of the game situation. In
Fig. 7 the three data paths are integrated in a third layer. At the third layer each neuron represents an
interaction between the typified movements of defense, offense and the ball. The last step consists of
applying a semantic meaning to the network – i.e., to each neuron – in the third layer. To this aim we
used a part of the manually categorized data as input to the architecture. After running through the
340 A. Grunz et al. / Human Movement Science 31 (2012) 334–343

Fig. 5. Using the sliding window technique the trajectory can be transformed into a set of vectors. For a fixed window length of
5 s we obtain one vector with five entries. Each entry is given by the corresponding cluster number of the activated neuron.

architecture neurons were activated on the third layer by sequences of labeled data, enabling a first
and still rather rough similarity-based automatic classification of these neurons. This first step of clas-
sification (what we call ‘‘coloring’’ because it results in a colored landscape of clusters) is the basis for
the semantic interpretation, where an expert examines each neuron to fix the semantic category of the
interaction prototype encoded in that neuron. If an encoded interaction doesn’t represent a meaning-
ful pattern, the corresponding neuron will be colored with a special color for meaningless interactions.
Instead of training and testing one network with weight vectors containing the four defense and four
offense players and the ball we apply a hierarchical architecture. This approach has the advantage that
 A. Grunz et al. / Human Movement Science 31 (2012) 334–343 341

Fig. 6. Illustration of the sliding window technique: The input sequence is decomposed into a set of fixed sequences. Each
sequence has a length of 5 s and represents a movement pattern. Each number encodes a different constellation type.

Fig. 7. Schematic illustration of the architecture of artificial neural networks used to classify different patterns and variations in
patterns. Vertically three blocks are given: On the left side the defense constellations are handled. In the middle the offense
constellations are handled. On the right side the ball is handled. Horizontally three layers can be distinguished: In the top layer
we typify constellation. In the middle layer movements of constellations are typified. The bottom layer typifies the interaction
of the typified movement of the defense, the offense and the ball. In the third layer activated neurons are recorded in a game log.

the complete information of a tactical pattern is reduced in several small steps resulting in better
topology preservation.

1.6. Implication: sliding window technique

One major problem that has been mentioned above but yet not discussed in detail is that of trans-
ferring trajectory information from the first network layer to neuron information – i.e., weight vectors
342 A. Grunz et al. / Human Movement Science 31 (2012) 334–343

– of the second layer: As stated before, information sequences for training and testing neurons have to
have a fixed length because the weight vectors of network neurons are defined this way. But trajec-
tories as mappings of movement patterns normally do not have such fixed lengths. Measured in units
of time, rather often we have patterns of a length of 5 s, but longer and shorter patterns are possible.
There are two ways to handle this problem. On the one hand different exemplars of networks can
be used for movement patterns of different length, e.g., one network for patterns of length 3, one for
length 4, one for length 5, and so on. This approach needs a huge amount of training data for the dif-
ferent networks, which moreover makes it difficult to compare sequences of the same type but of dif-
ferent length.
On the other hand the architecture can be restricted to a fixed sequence length – e.g., correspond-
ing to the smallest detectible pattern length. In this case a pre-processing and a post-processing step
are necessary. The pre-processing uses the above mentioned sliding window technique for stepwise
departing trajectories in pieces of equal length, which are classified by the neurons of the second-
level-networks. After that classification, the post-processing can compose larger patterns from the
fixed length entries of the classified neurons. However, this procedure is not easy because it has to
be decided automatically which partial patterns can be or have to be combined to complete ones,
preserving the original semantic meaning. Nevertheless, based on rather good experiences with the
sliding window method, we are following the fixed length approach.

2. Results and discussion

Our method was evaluated on the positional data, which have been drawn from the soccer world
championship final 2006. Restricting the investigation to game initiations we will show that the archi-
tecture is capable of classifying different patterns and variations in a pattern. We simplify the inves-
tigation by reducing the interaction to two different cases: on the one hand we consider the
interaction between the four Italian defense players, the four French attack players, and the ball. On
the other hand we consider the interaction between the four French defense players, the four Italian
attack players and the ball. During a game opening other players could be involved too as the number
of involved players isn’t usually fixed. Therefore some game initiations won’t be found.
During the soccer world championship final 2006 positional data had been recorded 25 times per
second resulting in 165.325 datasets. Using only one dataset per second the data basis for training was
reduced to 6613 datasets. Each dataset contained all 22 player positions and the ball position. For each
network on the first layer we constructed the needed input vectors, which were fed into at random.
Successive networks were trained using the output of the previous layer. 2228 datasets labeled as
short game initiation and 235 datasets labeled as long game initiation were extracted from the manual
categorization of the game by an expert. These datasets were used to color the network on the third
layer. Then an expert colored the remaining undefined neurons. In the productive phase the architec-
ture was fed with the original data sequence. On the third layer activated neurons were recorded in a
game log. The proposed post-processing step concatenates two sequences following in time course if
they belong to the same interaction type. For instance two sequences classified as short game initia-
tion and following each other on time course are concatenated by post-processing to a longer short
game initiation. After post-processing 110 game initiations remained in the game log. The manual cat-
egorization had found 131 game initiations. Thus our approach detected 84% of all game initiations.
We could demonstrate that our architecture of neural networks is basically capable of detecting
categories of tactical patterns. Preliminary results showed good classification accuracy. Our approach
has the advantage that each neuron has a clear meaning which is comprehensible at all times. The
main drawback results from using the sliding window technique. While this problem could be ade-
quately handled, the solution remains dissatisfying. Further work is necessary to find a better solution.
The same holds for the varying number of involved players which was kept constant in our
investigation.
Future investigations will focus on the detection of rare but relevant and meaningful tactical pat-
terns in the position data. Thus, we will try to explore creative tactical solutions in more depth
(Memmert, in press). Another application of SOM is the simulation of creative actions in sport games.
 A. Grunz et al. / Human Movement Science 31 (2012) 334–343 343

Mapped to SOM, this means that neurons should have the ability to represent not only frequent but
also and in particular seldom actions. If such a neural network is calibrated with respect to success,
then the time series of a process is mapped to a trajectory, where the neurons can be recognized to
correspond to creative actions (Memmert & Perl, 2009b).
Nevertheless, the final aim of the SOM approach is to reduce time and effort for the transfer from
recorded data to tactical information: By experience, it will take about 5 h per soccer game to manu-
ally observe and categorize the game video and fix the main tactical events. Net-based support of this
process can drastically reduce the required effort and time so that a tactic analysis could even be done
during the halftime break of a soccer game.

References

Bishop, C. M. (2007). Pattern recognition and machine learning. Berlin: Springer.

Grunz, A., Memmert, D., & Perl, J. (2009). Analysis and simulation of actions in games by means of special self-organizing maps.
International Journal of Computer Science in Sport, 8, 22–37.
Hughes, M., & Franks, I. M. (Eds.). (2009). National analysis of sport. Second Edition. Systems for better coaching and performance in
sport. London: Routledge.
Jäger, J., Perl, J., & Schöllhorn, W. (2007). Analysis of players’ configurations by means of artificial neural networks. International
Journal of Performance Analysis of Sport, 3, 90–103.
Kohonen, T. (1995). Self-organizing maps. New-York: Springer.
Memmert, D., & Perl, J. (2006). Analysis of game creativity development by means of continuously learning neural networks. In
E. F. Moritz & S. Haake (Eds.). The Engineering of Sport 6 (Vol. 3, pp. 261–266). New York: Springer.
Memmert, D., & Perl, J. (2009a). Game creativity analysis by means of neural networks. Journal of Sport Science, 27, 139–149.
Memmert, D., & Perl, J. (2009b). Analysis and simulation of creativity learning by means of artificial neural networks. Human
Movement Science, 28, 263–282.
Memmert, D. (in press). Sports and creativity. M. Runco & S. Pritzker (Eds.), Encyclopedia of creativity, 2nd Edition. Elsevier.
Perl, J. (2002). Adaptation, antagonism, and system dynamics. In G. Ghent, D. Kluka, & D. Jones (Eds.), Perspectives – The
multidisciplinary series of physical education and sport science, 4 (pp. 105–125). Oxford: Meyer & Meyer Sport.
Perl, J. (2004). A neural network approach to movement pattern analysis. Human Movement Science, 23, 605–620.
Perl, J. (2008). Neural network-based process analysis in sport. In J. R. R. Dopico, J. Dorado de la Calle & A. P. Sierra (Eds.),
Encyclopedia of artificial intelligence, III, (pp. 1212–1218).
Perl, J., Memmert, D., Bischof, J., & Gerharz, Ch. (2006). On a first attempt to modelling creativity learning by means of artificial
neural networks. International Journal of Computer Science in Sport, 5, 33–37.
Pfeiffer, M., & Perl, J. (2006). Analysis of tactical structures in team handball by means of artificial neural networks. International
Journal of Computer Science in Sport, 5, 4–14.
Reilly, T., Cabri, J., & Araujo, D. (Eds.). (2005). Science and Football V. The Proceedings of the Fifth World Congress on Science and
Football. London: Routledge.