Ball-Receiving Skill Dependent on Centering in
Soccer Simulation Games
Kazuaki Maeda
Akinori Kohketsu
Tomoichi Takahashi
kaz@solan.chubu.ac.jp g94117@isc.chubu.ac.jp ttaka@isc.chubu.ac.jp
Department of Business Administration and Information Science,
Chubu University
1200 Matsumoto, Kasugai, Aichi 487-8501, JAPAN
Abstract. This paper describes an e ective ball-receiving skill. When
soccer games are played in real life, players generally must make consecutive actions in one play, for example, running, receiving, and shooting
a ball. We believe that the same is true in the case of simulation soccer
games. Therefore, we designed an experiment to check how changing ballreceiving methods which is dependent on the centering patterns in uence
scoring goals. The experiment shows that one ball-receiving method is
more e ective than the others. The result is embedded into our soccer
team, Kasugai-bito II, and the e ectiveness is discussed in games.
1
Introduction
In the last few years, many researchers have devoted their e orts toward RoboCup
(The World Cup Robot Soccer)[2, 3, 4]. In RoboCup, robots or agents play a soccer game under given constraints. RoboCup was proposed as a standard problem
to promote AI and intelligent robotics research. It consists of three competition
tracks: Simulator League, Real Robot Small Size League, and Real Robot Middle
Size Leage[1]. We have investigated only Simulator League.
RoboCup has features di erent from typical traditional AI problems such
as computer chess. Features of computer chess are static and non-real-time. A
computer with powerful and large parallel processors, enormous knowledge, and
sophisticated software, can nd the most e ective move in a few minutes, and it
defeated the world chess champion [6].
For RoboCup, however, situations change dynamically and in real-time. Therefore, we cannot apply the traditional techniques to RoboCup. To develop a high
level robot soccer game, we must investigate intelligence (e.g. learning) for individual agents and teams.
The purpose of this paper is to con rm the e ectiveness in changing the
soccer agents' action dependent on situations. At this time, much of the work
about intelligent agents is in the planning phase. However, it is important to
con rm the e ectiveness before we implement our intelligent agents. Our current
soccer agents, Kasuga-bito II, do not have intelligent ability, but are hand-coded
to change ball-receiving methods which are dependent on centering. After this
con rmation of e ectiveness, we will implement the new intelligent agents. If
M. Asada and H. Kitano (Eds.): RoboCup-98, LNAI 1604, pp. 152-161, 1999.
c Springer-Verlag Heidelberg Berlin 1999
Ball-Receiving Skill Dependent on Centering in Soccer Simulation Games
153
new intelligent agents are the same level as our hand-coded agents, they will be
regarded as having good skill.
In section 2, we de ne centering patterns and ball-receiving methods. In
section 3, we explain the outline of our experiment and evaluate the e ectiveness
for the changing of receiving methods. In section 4, we describe the results of
the match games where we applied our skill. Finally, we summarize the paper.
2
Centering Patterns and Receiving Methods
2.1 Real Soccer Games
From the experience of playing soccer games in real life, players generally must
make consecutive actions in one play, for example, running, receiving, and shooting a ball. In the consecutive actions, they must receive the ball by the most
e ective methods. That is, we should select one method to receive the ball. For
example, when the centering ball moves toward the goal, we usually turn toward
the goal and receive the ball, then shoot it. Another example is when the centering ball lands outside the goal. In this case, we usually turn back against the
goal and receive the ball, then shoot it. Generally speaking, we decide the next
action by considering current situations.
When human players shoot the ball toward the goal, it is said that as the number of times the ball is touched increases, the shooting success rate decreases[5].
Moreover, it is said that direct or one-trapping shots occupy 70{80% of the successful shoots. This points out the need for real-time processing for the soccer
agents. To satisfy this need, the agents should shoot the ball toward the goal as
soon as they receive it in the penalty area.
2.2 Classi cation of Centering Patterns
We can classify centering as the point of relationship between the player and the
ball. We call this the centering patterns.
We introduce 4 centering patterns such as depicted in Figure 1. The patterns
have di erent features relative to the position of the player (the \shooter") and
the direction of the ball.
Our classi cation is as follows:
1. When the X-position of the ball is the same as the X-position of the shooter,
the Y-position of the ball is
{ positive to the Y-position of the shooter (1 and 3 in Figure 1), or
{ negative to the Y-position of the shooter (2 and 4 in Figure 1).
2. The gradient of the ball-trajectory is
{ positive (3 and 4 in Figure 1), or
{ negative (1 and 2 in Figure 1).
These combinations make 4 centering patterns. To simplify our discussion, we
will call the patterns Centering-1, Centering-2, Centering-3, and Centering-4.
154
Kazuaki Maeda et al.
Fig. 1.
2.3
Centering Patterns
Prediction and Receiving Methods
Prediction of the Ball Position: For receiving the ball, we need to design a
prediction of ball movement by using see-information from the soccer server. We
will describe our prediction in this section.
When we know the current ball position B(t) at time t, we can calculate the
predicted ball position B(t + 100ms) at 100ms later below:
1. Calculate the distance, 1(t), of the ball movement by using changes of relative distance and angle between the ball and the player.
1(t) = (t) 0 (t 0 100ms)
B
B
2. Multiply 1(t) by the default Ball-Decay 0.96 and add the current position
B(t) to it.
(t + 100ms) = B(t) + 0:96 1(t)
B
When we predict the ball position at more than 100ms later, we can obtain it
by repeating the above calculation.
From the experience of soccer games in real life, we have
found some receiving methods from the point of relationship between the player
and the ball. This naturally led us to introduce 3 methods for receiving the ball.
The methods are di erent based on the relative position between the shooter
and the ball such as depicted in Figure 2.
At the time t, we describe the shooter position as S(t) and the ball position
as B(t). Moreover, we describe t1 as the moment when the shooter reaches on
the ball-trajectory (maybe the predicted ball-trajectory), and t2 as the moment
when the shooter receives the ball. This implies S(t2) = B(t2). To simplify our
discussion, we give the name to each method below:
Receiving Methods:
Ball-Receiving Skill Dependent on Centering in Soccer Simulation Games
155
Shooter
Ball
Fig. 2. Backward-Receiving, Simultaneous-Receiving, and Forward-Receiving
This method is used when the shooter reaches the predicted ball position in advance of the ball. The feature is
Backward-Receiving
S(
t2
t1) = B(t2)
= t1 + 300ms
That is, the shooter waits for the ball before receiving it. Therefore, he
can modify his position for the unpredictable ball movement. This method,
however, has a drawback in that there is a delay in waiting for the ball.
Simultaneous-Receiving This method is used when the shooter receives the
ball as soon as he reaches the predicted ball position. The feature is
t2 = t1
This method is the fastest among three methods. However, the shooter usually turns his back against the goal to receive the ball. Therefore, this method
has a drawback in that he cannot see the situation around the goal.
Forward-Receiving This method is used when the shooter runs after the ball
and receives it. The feature is
S(
t1) = B(t1 0 100ms)
The shooter can easily see the situation around the goal (especially the
goalkeeper). He utilizes his see-information around the goal and can shoot
the ball e ectively. This method, however, has a drawback of sometimes
having to wait before receiving the ball.
These receiving methods make one action in consecutive actions to score.
The next action is to shoot the ball. From the experience of soccer games in real
life, our agents shoot the ball toward the goal as soon as they receive it[5].
156
3
Kazuaki Maeda et al.
An Experiment on Receiving and Shooting
We believe that soccer agent's skill of their shooting is related to their scoring
ability. Therefore, we designed a skill experiment and examined the scores to
evaluate our agents. If our well-designed agents improve their ability to score,
we can conclude that the skill of our agents has improved.
We made an experiment to try 3 receiving methods from 4 centering patterns
respectively. Figure 3 depicts an overview of the experiment. We describe here
the experiment and the result.
Fig. 3.
3.1
An Overview of the Experiment
Fig. 4.
Positions of the Players
The Condition of the Experiment
We designed the experiment with the following conditions:
{
{
{
We prepared two players for o ense. One kicked a ball according to a centering pattern. We called him the centering-player. The other was the shooter.
We also prepared one goalkeeper for defense.
We made the experiment in the coaching mode on the soccer server (version
3.17) to position the players and the ball.
The shooter was positioned at random within a range of
X-position: 33 37 and
Y-position: 020 010.
The centering-player was positioned at the place where the direction from
the shooter was
150 degrees for Centering-1 and Centering-2, and
-150 degrees for Centering-3 and Centering-4.
Moreover, the distance between the centering-player and the shooter had a
range of 63 around 17.5. The goalkeeper was positioned at 2.5 away from
the goal. Figure 4 depicts the positions of players.
Ball-Receiving Skill Dependent on Centering in Soccer Simulation Games
{
{
{
{
157
The beginning position of the ball was at the centering-player's feet.
The centering-player kicked the ball to the shooter by kick-power 100 toward
the relative direction of
5 20 degrees for Centering-1 and Centering-3, and
05 020 degrees for Centering-2 and Centering-4.
The shooter shot the ball as soon as he received it.
If it took less than 6 seconds to score the goal after the centering-player
kicked the ball, the trial was regarded as success. Otherwise, the trial was
regarded as a failure.
The number of trials was as follows:
{
{
The centering-player kicked the ball to the shooter 300 times for each centering pattern. There were 4 centering patterns. Therefore, he kicked the ball
1200 times.
The shooter received the ball 100 times by using each method for each centering pattern.
We adopted the goalkeeper of AT-Humboldt team (Germany) which won the
championship at the 1st RoboCup Worldcup Simulation League held on 1997.
3.2
Experiment Result
The results of the experiment are shown in Table 1. The table shows the success
rates of shooting the ball to 4 centering patterns by 3 receiving methods. The
underlined rates are the most e ective receiving methods for each centering
pattern. We consider here the result for each centering pattern.
Success Rate of Each Receiving Method (%)
1 2 3 4
Backward-Receiving 42 40 39 62
Simultaneous-Receiving 63 46 66 50
Forward-Receiving 33 53 48 28
Table 1.
First, the table shows that the Simultaneous-Receiving is the most e ective
method for Centering-1. By comparing this with soccer games in real life, it is
easily understood.
Secondly, the table shows that Forward-Receiving is the most e ective method
for Centering-2. In a real soccer game, however, we may nd that BackwardReceiving is the most e ective method. This is mainly because the shooter,
when Forward-Receiving, can see the situation around the goal.
Thirdly, the table shows that Simultaneous-Receiving is the most e ective
method for Centering-3. In a real soccer game, however, we may nd that
158
Kazuaki Maeda et al.
Forward-Receiving is the most e ective method. This is mainly because it is likely
that the goalkeeper will catch the ball if the shooter uses Forward-Receiving.
Finally, the table shows that Backward-Receiving is the most e ective method
for Centering-4. In a real soccer game, however, we may nd that ForwardReceiving is the most e ective method. This reason is the same as one given for
Centering-3.
3.3
Statistical Testing
To examine the experiment results statistically, we tried hypothesis-testing. At
rst, we set a hypothesis:
The receiving methods are not dependent on the success rate to score
the goal.
Next, we made a 2 2 3 contingency table, where
f
f
g2
g . Then we
shoot success, failure
Backward-Receiving, Simultaneous-Receiving, Forward-Receiving
tested it, where the signi cance level is 1% and the rejection region is more than
9.28. Table 2 shows that our hypothesis was rejected to Centering-1, Centering-3,
and Centering-4. That is, the receiving methods are dependent on the centering
patterns except in the case of Centering-2.
Table 2. Result of Statistical Testing
1
2
3
4
testing reject accept reject reject
19.08
4
2.79
23.89 15.13
Application to a Soccer Game
As mentioned above, changing the receiving methods which is dependent on
centering patterns is usually e ective to score the goal. We will con rm here the
e ectiveness in a soccer simulation match game.
First, we prepared two teams for comparison:
All players select one of three receiving methods at random.
All players select the most e ective method dependent on the centering
patterns.
Rand
Advc
For both teams, the receiving methods are only used for the players which are
positioned within 30 meters from the center of the goal (52.5, 0) such as depicted
in Figure 5.
Ball-Receiving Skill Dependent on Centering in Soccer Simulation Games
Fig. 5.
159
Region to Change the Receiving Methods
We selected two opposing teams. One was AT-Humboldt team. The other was
Andhill team (Tokyo Institute of Technology, Japan). Andhill won the second
prize in the 1st RoboCup Worldcup Simulation League. The combinations were
4 match games. We tried each match game for 100 minutes1 .
The di erences in the scores are presented in Table 3. When we changed
teams from Rand to Advc, the di erence in the scores was changed for both
AT-Humboldt and Andhill. These results show the e ectiveness of our changing
methods in a soccer simulation match game.
The Di erences in the Scores
KasugaAT
Di erence Kasuga- Andhill Di erence
bito II -Humboldt
bito II
Rand 31
33
-2
28
37
-9
Advc 36
31
+5
33
35
-2
Table 3.
5
Conclusion
The purpose of this paper was to examine the e ectiveness of changing the ballreceiving methods in RoboCup simulation soccer game. For that purpose, we
made an experiment to discover the most e ective receiving methods which are
dependent on the centering patterns. As a result, we found e ective receiving
methods except in the case of one centering pattern.
Acknowledgements
We wish to thank developers of AT-Humboldt and Andhill for freely distributing
their soccer clients. We also thank the anonymous reviewers for some suggestions.
1
We manipulated the length of games using the con guration le of soccer-server
160
Kazuaki Maeda et al.
References
1. The RoboCup
Federation.
RoboCup:
The
Robot
World
Cup
Initiative.
http://www.robocup.org, 1998.
2. Hiroaki Kitano, Minoru Asada, et al. The RoboCup: The Robot World Cup Initiative. In Proceedings of IJCAI-95 Workshop on Entertainment and AI/Alife, pages
19{24, 1995.
3. Hiroaki Kitano, Minoru Asada, et al. RoboCup: The Robot World Cup Initiative.
In Proceedings of First International Conference on Autonomous Agent, 1997.
4. Hiroaki Kitano, Milind Tambe, et al. The RoboCup Synthetic Agent Challenge 97.
In Proceedings of IJCAI-97, 1997.
5. Jiro Ohashi, Kozo Tajima, and Takashi Kakemizu. Science of Soccer (in Japanese).
Tokyo Denki University Press, 1997.
6. Herbert A. Simon and Toshinori Munakata. AI Lessons. Communications of the
ACM, 40(8):23{25, 1997.
Appendix: Team Description
Our team, Kasugabito-II, was runner-up in the Japan Open 98 Simulation
League. The most important feature of Kasugabito-II is, we mentioned it in
this paper, changing ball-receiving methods which is dependent on the centering
patterns. We describe other features here.
Formation and Roles
Kasugabito-II is composed of 3 Defenders (DFs), 5 Mid elders (MFs), 2 Forwards
(FWs) and 1 Goalkeeper (GK) such as depicted in Fig.6.
DFs defend away a little bit from their goal to trap in o -side. MF*s in Fig.6
sometimes participate in an attack by dribbling. FWs stand nearby the penalty
area and wait for centering from a MF.
In Kasugabito-II, each player has the same algorithm for predicting the ball
position (see section 2.3). About the movement, however, each player's action is
dependent on the position. For instance, a GK and DFs wait for the ball until
the ball enters within a distance, and players move to receive it. This is because
they should receive the ball safely by using more reliable prediction of the ball
position. MFs and FWs move earlier toward the ball to rival opposite players to
receive the ball.
Implementation
Kasugabito-II is written in C and it consists of about 3,000 lines. We are developing it on multiple SPARCstations which are running under Solaris 2.5.
We were using two workstations (SPARCstation 5/110, memory 64M bytes)
in the early development. In the recent development, we are using ve workstations;
Ball-Receiving Skill Dependent on Centering in Soccer Simulation Games
FW
FW
MF*
MF*
MF
MF
DF
MF
DF
DF
GK
Fig. 6. Formation
{
{
One Sun Ultra 1/140 (Memory 96M bytes)2
Four Sun Ultra 1/170 (Memory 128M bytes)3
We would like to thank people for lending these workstations to us.
2
3
This machine is supported by Nippon Steel, Inc. and Nissho Electronics, Inc.
These machines are supported by Nihon Sun Microsystems, Inc.
161