PREDICTING PERFORMANCE IN TEAM GAMES

The Automatic Coach

Guillermo Jim´enez-D´ıaz

, H´ector D. Men´endez

, David Camacho

and Pedro A. Gonz´alez-Calero

Complutense University of Madrid, Computer Science Department, Madrid, Spain

Universidad Aut´onoma de Madrid, Computer Science Department, Madrid, Spain

Keywords:

Videogames, Clustering.

Abstract:

A wide range of modern videogames involves a number of players collaborating to obtain a common goal.

The way the players are teamed up is usually based on a measure of performance that makes players with a

similar level of performance play together. We propose a novel technique based on clustering over observed

behaviour in the game that seeks to exploit the particular way of playing of every player to ﬁnd other players

with a gameplay such that in combination will constitute a good team, in a similar way to a human coach.

This paper describes the preliminary results using these techniques for the characterization of player and team

behaviours. Experiments are performed in the domain of Soccerbots.

1 INTRODUCTION

Online games are the fastest growing market in en-

tertainment. A key feature that explains online game

success is their social nature. In these games, players

collaborate and compete against other players around

the world. An online multiplayer match can be ar-

ranged basically in two ways, either a group of friends

conﬁgure their own match to play together, or an in-

dividual player joins a match from a list of matches

currently in progress in the server.

Most online multiplayer games offer an entrance

hub for matchmaking. The matchmaking process

helps individual players to ﬁnd a match where they

are likely to have an enjoyable gameplay experience.

The most sophisticated matchmaking services pro-

vide recommendations based on the player skill level.

The best known skill rating system is the Elo system

(Elo, 1978), adopted by the World Chess Federation

to rank chess players, which is also used as a rat-

ing system for multiplayer competition in a number

of computer games. The TrueSkill

rating system

(Herbrich et al., 2006) is an evolution of the Elo sys-

tem and it is employed in the Xbox Live online gam-

ing system.

Matchmaking based on the players skill level as-

sumes that games have to be balanced in order to be

fun to play. However, this assumption is clearly re-

stricted for multiplayer games, where different play-

ers usually adopt different roles in the team, and per-

formance as well as gameplay satisfaction depends on

the careful combination of roles. Just imagine a soc-

cer team with only goalkeepers.

We propose to extend matchmakingin multiplayer

online games with information about the particular

abilities of the players that make them more apt for

playing a given role in the team. For such a match-

making approach it should be possible to infer the best

role for a player based on his observed behaviour in

the game. This paper demonstrates that it is possible

to do using classiﬁcation and clustering techniques.

The rest of the paper runs as follows. Next Sec-

tion brieﬂy introduces the classiﬁcation and clustering

techniques used in the paper. Section 3 describes the

experimental setup in the Soccerbots domain, while

Section 4 presents the experimental results. Finally, in

Section 5, conclusions and future work are discussed.

2 CLUSTERING BEHAVIOURS

There has been used different approximations based

on Machine Learning and Data Mining to model the

player behaviour. These approaches, usually named

human or robot behaviour modelling, has been ap-

plied in different domains like Robossocer simula-

tions (Grollman and Jenkins, 2007; Aler et al., 2009;

401

Jiménez-Díaz G., D. Menéndez H., Camacho D. and A. González-Calero P..

PREDICTING PERFORMANCE IN TEAM GAMES - The Automatic Coach.

DOI: 10.5220/0003185104010406

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 401-406

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Leng et al., 2010). In this work some popular meth-

ods from Data Mining (Larose, 2005), such as clus-

tering or classiﬁers algorithms, are combined to ex-

tract the behaviour of players during a game session,

and to classify how they are is playing against their

opponents. The proposed techniques are exempliﬁed

using simulations of soccer matches with robots, so

the techniques are employed to characterize the be-

haviour of a team and its individual robots during a

soccer match.

To apply Data mining methods it is usually nec-

essary to make an intensive phase of data preprocess-

ing. Initially the information must be analysed and

stored in some kind of database system, cleaned and

separated. This preprocessing phase is used to avoid

outliers, missclassiﬁcations and missing data. The

second step is related to normalization. It allows to

compare data features with different kind or range of

values. Z-Score (Carroll and Carroll, 2002) and Min-

Max (Han and Kamber, 2006) normalization methods

are commonly used for preprocessing the data.

Both normalization algorithms takes the attribute

records and they try to ﬁnd a standard range for them.

Min-Max has a ﬁxed range, [0,1], while Z-Score de-

pends of mean and standard deviation. These algo-

rithms obtain the normalized values from data using

the following equations:

• Min-max: It computes maximum and minimum

values of the attributes and it applies the equation:

′

x− min(X)

max(X) − min(X)

• Z-Score: It computes mean and standard devia-

tion of the values and it applies the equation:

′

x− mean(X)

SD(X)

Once data is preprocessed and normalized, a high

number of possible data mining methods could be

used to analyse it. We have selected some popular

Classiﬁcation and Clustering algorithms. C4.5 (Quin-

lan, 1993) and CART (Breiman and Stone, 1984)have

been employed to extract information from different

datasets. We have employed these methods to ana-

lyze the strategies of the whole teams using the match

results, i.e. comparing the time in a particular zone,

the number of kicks or the distance to the center of

the ﬁeld for each player. These algorithms allow to

automatically compare (classify) how the teams are

playing using the individual results from each player.

The main problem with these algorithms is the length

of trees generated. C4.5 and CART algorithms are not

able to directly process the huge amount of data (see

Section 3), so a binning method (Larose, 2005) has

been used to control the tree’s size.

Finally, clustering methods have been used to

classify what kind of behaviour (offensive, defensive,

or balanced) is deployed by a particular team in a

match. Simple algorithms such as statistical algo-

rithms (Expectation Maximization, EM) or K-means

have been tested to compare the available matches.

With these methods it is possible to establish the real

behaviour of a team and then automatically classify

them using their behaviour.

2.1 A Short Description on Clustering

and Classiﬁcation Algorithms

From the wide area of clustering and classiﬁcation

methods, only several well known algorithms have

been selected to evaluate how these techniques can be

used to automatically extract the behaviour of teams

and the team members. From classiﬁcation methods,

some popular approaches are based on decision trees.

In these algorithms a tree is built from scratch using a

set of values that characterizes a feature. These trees

are usually employed to extract a set of rules that later

can be used to classify a new pattern. The algorithms

used in this work to compare the results obtained by a

particular team using their results in a gameplay are:

• C4.5 (Quinlan, 1993) is an extension of ID3 al-

gorithm (Quinlan, 1986). Both algorithms use the

concept of information entropy reduction. C4.5

has the main following characteristics: it produces

a tree of more than one variable shape; for cat-

egorical attributes it produces a separated branch

for each value of the categorical attribute; and it

uses entropy reduction methods to select the opti-

mal split. The entropy reduction method consists

on (Larose, 2005):

– Split the records into subsets by choosing a in-

stance or candidate.

– The mean information of each subset is calcu-

lated as the weighted sum of the entropies for

the individual subsets.

– Once the mean information is computed for all

the candidates, the information gain for each of

them is calculated, it is based on the entropy.

– The candidate split that has the greatest infor-

mation gain is chosen.

– Finally, the process continues recursively until

all the instances are classiﬁed.

• CART (Breiman and Stone, 1984) is another pop-

ular algorithm that builds a binary tree using op-

timal splits. These optimal splits are chosen ac-

cording to a criteria based on a measure of the

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

402

“goodness” of a candidate split at a node. This

measure is calculated using the neighbor nodes

and conditional probability on the records con-

ditionated to these nodes. The optimal split is

whichever split that maximizes this measure over

all possible splits at the selected node. Recur-

sively, CART splits the records in the training data

set into subset of records with similar values for

the target attribute (Larose, 2005).

Finally, the clustering algorithms used to classify

the real observed behaviour in a set of gameplays (we

call them the team behaviour) are:

• Expectation-Maximitation (Robert Hogg and

Craig, 2005), or simply EM, is a blind clustering

algorithm that tries to classify and create the clus-

ters for the data. This algorithm is useful when

data is hidden or missed. Initially, it takes a like-

lihood and tries to maximize it. The process con-

sists on apply the two following steps iteratively

until it ﬁnishes:

– Expectation step: Calculate the expected value

of the log-likelihood function and redeﬁne it.

– Maximization step: Find the parameter that

maximizes the likelihood function.

• K-means (MacKay, 2003) is another popular and

well knownalgorithm. It is a straightforward clus-

tering guided method (usually by a heuristic or di-

rectly by a human) to try to classify data in a ﬁxed

number of clusters. The number of clusters can

be predeﬁned or it can be estimated using heuris-

tics or other kind of algorithms, like genetic algo-

rithms (Gonzalez-Pardo et al., 2010). This algo-

rithm runs in 5 steps (Larose, 2005):

– Deﬁne (ﬁx) the number of clusters (k).

– Assign k records to be the initial cluster center

location.

– For each record, ﬁnd the nearest cluster center.

– For each of the k clusters, ﬁnd the cluster cen-

troid, and update the location of each cluster

center to the new value of the centroid.

– Repeat the two last steps until a convergence

criteria or a termination condition is reached.

3 EXPERIMENTAL SETUP

The approaches described in this paper has been

tested using a dataset generated using Soccerbots.

Soccerbots

simulates the dynamics and dimensions

Soccerbots: http://www-2.cs.cmu.edu/∼trb/TeamBots/

Domains/SoccerBots/

of a regulation RoboCup small size robot league

game. Two teams of ﬁve robots compete in a soc-

cer ﬁeld by pushing and kicking a ball into the oppo-

nent’s goal. This simulator has been employed as a

sandbox in several works related to the application of

different machine learning techniques in multiagent

systems (Aler et al., 2009; Leng et al., 2010).

The data has been extracted using SBTourna-

ment

, a tool for generating Soccerbots tournaments

and trace generation of robot behaviour. SBTour-

nament extracts periodically the position, direction

and velocity for every robot and the ball during a

match. Additionally, the kick actions and goals are

asynchronously extracted. SBTournament uses these

traces to generate CSV ﬁles about every robot, team

and matches played. Finally, the dataset employed

in our evaluation has been enhanced computing some

statistical data extracted from the CSV ﬁles, described

below.

The dataset contains information about ∼ 15000

matches played by 74 different teams implemented by

students of Computer Science at Complutense Uni-

versity of Madrid during different academic courses.

There are three different types of information con-

tained in the dataset:

Information about each Robot. The dataset stores

statistics about every robot that has participated

in a match. For every match and every robot, we

have the number of goals scored and kicks per-

formed, the time the robot spent in its own ﬁeld

and opponent ﬁeld and in its own and in the op-

ponent goalkeeper area, the time the robot spent

in “ball possession” and the average distance be-

tween the robotand the ball, the center of the ﬁeld,

its own goal and the opponent goal.

Information about each Team during a Match.

The information about each robot is compiled

for generating the global statistics for each team

during the match. These team statistics contain

the aggregation and the average values from

every feature extracted from the robots that make

up the team, such as the number of goals scored

and received by the team, the sum up of the team

robot kicks, the average time that the team robots

spent in their own ﬁeld and in the opponent ﬁeld

or the total time that the team robots spent in “ball

possession”, among others.

Global Information about every Team. Using the

information about a team during all the played

matches we generate a set of descriptive statistics

that summarizes the global team behaviour. These

SBTournament: http://gaia.fdi.ucm.es/projects/

soccerBots/SBTournament 1.2.zip

PREDICTING PERFORMANCE IN TEAM GAMES - The Automatic Coach

403

statistics contain information about the number of

goals scored and received by the team, the number

of wins, lost and ties or the total number of kicks

performed, among others.

For testing our clustering methods we have to

choose the best soccer teams, as we will describe in

Section 4.3. These teams have been selected from the

winners of the tournaments

. celebrated every year at

the Complutense Universityof Madrid with the robots

implemented by the students –the robots employed to

create the dataset.

4 RESULTS

The data previously described has been processed fol-

lowing the steps described in the next subsections in

order to allow their automatic analysis.

4.1 Preprocessing

The CSV ﬁles described in Section 3 were separated

in two groups: individual player results and team re-

sults. These data are used to analyze how the player

behaviour inﬂuences on the team.

Owing to the fact that the program can be conﬁg-

ured with different match time, the data needs to be

normalized (see Section 2). A min-max standardiza-

tion is the easiest way to join this data and it simpli-

ﬁes the binning methods, although other normaliza-

tion method like Z-Score was used. Some characteris-

tics in the data (like different time duration in matches

played) affect negatively to mean and standard devia-

tion in the normalization process.

4.2 Exploratory Data Analysis

4.2.1 The Goals-kicks Results

The ﬁrst approximation of the data analysis reveals

that there are few strategies with satisfactory results.

These strategies are divided on defensive, balanced

and offensive. It is worth noting that a player gets

better scores when it drags the ball instead of kick it.

4.2.2 Type of Player

The exploratory data analysis reveals that the play-

ers have different behaviours that can be related be-

tween them. Choosing these special situations, the

data shows that a few of players have special features.

These players are categorized in:

The ﬁnal results of these tournaments are available at

http://gaia.fdi.ucm.es/grupo/software.html

• Keeper. This player is always close to the goal

area. It also kicks frequently the ball. The pattern

is based on the attributes ‘time spend in own goal’,

‘time spend in own ﬁeld’, ‘distance to the ball’,

‘number of kicks’ and ‘distance to the own goal’.

• Defender. This player usually stands in its ﬁeld

side, except when an opponent is close to the cen-

ter or it takes the ball. The pattern is based on the

attributes ‘time spend in own ﬁeld’, ‘time poses-

sion’ and ‘distance to the center’.

• Midﬁeld player. This kind of player has a ran-

dom behaviour. Usually, the whole team is ini-

tially programmed to have only midﬁeld players.

The pattern is based on the attributes ‘time spend

in own ﬁeld’, ‘time spend in opponentﬁeld’, ‘time

posession’ and ‘center distance’.

• Kicker-Forward. This kind of player usually

stands in the opponent ﬁeld side, close to the cen-

ter. When it takes the ball, it tries to kick it.

The pattern is based on the attributes ‘time spend

in opponent ﬁeld’, ‘time posession’, ‘center dis-

tance’, ‘number of kicks’ and ‘opponent goal dis-

tance’.

• Dragger-Forward. This player also stands in the

opponent ﬁeld side, but when it takes the ball, it

drags the ball to the opponent goal. The pattern

is based on the attributes ‘time spend in opponent

ﬁeld’, ‘time posession’, ‘center distance’, ‘num-

ber of kicks’, ‘opponent goal distance’ and ’time

spend in the opponent goal’.

4.2.3 Type of Team

The team behaviour is similar to the player behaviour.

Following the same analysis criteria, the teams can be

divided into three different main categories:

• Defensive. This kind of team usually plays in its

ﬁeld side. It usually has a keeper and at least a

couple of defenders.

• Offensive. This kind of team usually plays closed

to the center. It has no keeper, and normally the

players do not spend too much time in its own

area.

• Balanced. A balanced team mixes the previous

behaviour. This kind of team is the most difﬁcult

one to be (automatically) discriminated because

the behaviour of the team can evolve from defen-

sive to offensive, or vice versa, during the match.

4.2.4 Decision Trees (C4.5/CART)

The initial analysis shows different strategies that the

teams have used. Therefore, it is important to classify

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

404

them with their results. Decision trees allow to ana-

lyze the result of the classiﬁcation showing rules that

could be generated by a good or a bad result based on

the initial data. Binning the data, the size of the tree

generated by CART and C4.5 and their accuracy are

shown in table 1.

Table 1: Leaves and accuracy percentage.

C4.5 CART

Goals Against 783 (87,7%) 121 (87,6%)

Goals For 2980 (57,6%) 277 (57,8%)

These results show that it is easier to classify a

bad strategy than a good one. This result demon-

strates that they have no similar behaviour to obtain

their goals. For this reason, clustering methods have

been selected to classify teams behaviour instead of

their strategy.

4.2.5 Clustering (EC/K-Means)

After the strategy classiﬁcation, the teams were clas-

siﬁed using the clustering methods. The results are

depicted in Table 2, which shows the results of sev-

eral variations in EM and K-Means algorithms with

team data. For every algorithm, it shows the number

of clusters generated and the percentage of instances

that belongs to the cluster (if there are more than one

cluster, the percentage of each cluster has been in-

dicated). First, Expectation-Maximization algorithm

has been tested with all the data and with the normal-

ized data. Finally, K-means classiﬁcation has been

tested choosing k values from 4 to 6 clusters.

Table 2: Clusters classiﬁcation. For each algorithm, the ta-

ble contains the number of clusters and the percentage of

instances classiﬁed in each cluster.

Algorithm Offensive Defensive Balanced

EM 1/30% 1/25% 2/33%,12%

EM (Nor.) 0 0 1 / 100%

4-Means 1/34% 1/25% 2/14%,27%

5-Means 1/32% 2/25%,13% 2/11%,20%

6-Means 2/27%,7% 2/23%,12% 2/11%,20%

Clustering shows that the teams have similar be-

haviours during the matches. The criteria of each be-

haviour has been described in previous section.

Clusters also show that there are teams that fol-

low multiple behaviours and teams that only follow an

speciﬁc behaviour. This result explains why it is dif-

ﬁcult to classify the balanced behaviour. 6-means al-

gorithm also classiﬁes correctly the team results. This

information helps us to ﬁnd a good opponent for each

team, and it will be used for the testing phase.

4.3 Tests

The best teams of the tournaments –described in Sec-

tion 3– have been classiﬁed using previous approach.

With the clusterings methods previously described the

results are used to deﬁne what kind of opponents are

“good” opponents against them.

4.3.1 Testing the Model

Table 3 contains the statistics about the global results

performed by the best teams. The best teams have

achieved a high number of winnings and ties but a

small number of looses.

Table 3: Best teams results.

Team Won Lost Tie

ISBCUnited 68,5% 1,75% 29,75%

JGDTeam 70,25% 3,5% 26,25%

WetTunaTeam 66,60% 5,25% 28,15%

The analysis results of the best teams behaviour

based on 6-means is shown in Table 4. This table

shows the percentage of team instances that belongs

to each cluster over the total instance number. These

tables show that the best teams have a defensive pro-

ﬁle while their offensive proﬁle is reduced. It could

be the reason because they also have a high percent-

age of tie games.

Table 4: Best teams classiﬁcation.

Team Offensive Defensive Balanced

ISBCUnited 3% 61,5% 35,5%

JGDTeam 11,11% 51,24% 37,65%

WetTunaTeam 7,98% 81,72% 10,3%

Tables 5 summarizes the behaviour of one of these

teams against its best opponents, selected for testing

the model. The table shows the rate of winnings and

their the classiﬁcation of 6-means cluster for these

teams. This simple analysis provides a good exam-

ple about how the proposed approach can be helpful

when it is necessary to ﬁnd a good team.

• Table 5 shows the teams that have defeated ISB-

CUnited. These results show that defensive teams

are usually more efﬁcient than offensive teams.

Nevertheless, completely and middle offensive

behaviours also obtain goods results.

PREDICTING PERFORMANCE IN TEAM GAMES - The Automatic Coach

405

Table 5: ISBCUnited best opponents.

Team Victories Offensive Defensive Balanced

MetalTeam 12,5% 50,01% 20,98% 29,01%

AmigosDeGattuso 12,5% 5,88% 93,28% 0,84%

D2JTeam 12,5% 10,49% 88,89% 0,62%

DJBTeam 12,5% 11,42% 83,33% 5,25%

JMPTeam 12,5% 100% 0% 0%

FourtyTwoTeamCBR 6,25% 43,75% 16,75% 39,5%

EspartanosRBR 6,25% 9% 83,75% 7,25%

5 CONCLUSIONS AND FUTURE

WORK

This paper has presented some preliminary results

about the use of classiﬁcation and clustering tech-

niques employed for characterizing the behaviour of

a player during a gameplay in order to ﬁnd interest-

ing players to play with. The techniques employed

and their results seem to be promising. They can be

applied to team-up human players or to ﬁnd compat-

ible bots to play with human players. We intend to

explore the application of these techniques in other

games with actual human data.

Our future work is related with the generation of

promising teams. We want to create a team with play-

ers of different teams and check how a player inﬂu-

ences in the composition of its team. This kind of

automatic human-like behaviours analysis could be

easily extended to other domains such as computer

(online) games or virtual worlds.

ACKNOWLEDGEMENTS

This work has been partly supported by: Span-

ish Ministry of Science and Education under grant

TIN2009-13692-C03-03, TIN2010-19872 and Span-

ish Ministry of Industry under grant TSI, 020110-

2009-205.

REFERENCES

Aler, R., Valls, J. M., Camacho, D., and Lopez, A. (2009).

Programming robosoccer agents by modeling human

behavior. Expert Systems with Applications, 36:1850–

1859.

Breiman, L., F. J. O. R. and Stone, C. (1984). Classiﬁca-

tion and Regression Tree. Wadsworth & Brooks/Cole

Advanced Books & Software.

Carroll, S. R. and Carroll, D. J. (2002). Statistics Made

Simple for School Leaders. Rowman & Littleﬁeld.

Elo, A. E. (1978). The rating of chess players: Past and

present. Arco Publishing, New York.

Gonzalez-Pardo, A., Granados, A., Camacho, D., and Ro-

drguez, F. D. (2010). Inﬂuence of music representa-

tion on compression-based clustering. In IEEE World

Congress on Computational Intelligence, pages 2988

– 2995. IEEE.

Grollman, D. H. and Jenkins, O. C. (2007). Learning

robot soccer skills from demonstration. In Inter-

national Conference on Development and Learning,

pages 276–281.

Han, J. and Kamber, M. (2006). Data mining: concepts and

techniques. Morgan Kaufmann.

Herbrich, R., Minka, T., and Graepel, T. (2006).

Trueskill

: A bayesian skill rating system. In

Sch¨olkopf, B., Platt, J. C., and Hoffman, T., edi-

tors, Advances in Neural Information Processing Sys-

tems 19, Proceedings of the Twentieth Annual Con-

ference on Neural Information Processing Systems,

pages 569–576. MIT Press.

Larose, D. T. (2005). Discovering Knowledge in Data. John

Wiley & Sons.

Leng, J., Fyfe, C., and Jain, L. (2010). Reinforcement

learning of competitive skills with soccer agents. In

Knowledge-Based Intelligent Information and Engi-

neering Systems, volume LNCS 4692, pages 572–579.

Springer.

MacKay, D. (2003). Information Theory, Inference and

Learning Algorithms. Cambridge University Press.

Quinlan, J. R. (1993). C4.5: Programs for Machine Learn-

ing. Morgan Kaufmann Publishers.

Quinlan, R. J. (1986). Induction of decision trees. Machine

Learning, 1:81–106.

Robert Hogg, J. M. and Craig, A. (2005). Introduction to

Mathematical Statistics. Pearson Prentice Hall.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

406