Could You Surpass Coach Troussier:
Controlling Interaction Among Multiple Agents
1. If You Were Coach Troussier
Recently, many have been voicing the opinion that the Japanese National Men's
Soccer Team has become stronger under the leadership of Coach Philippe Troussier.
Considering that the Japanese soccer team reached the top eight in the Sydney
Olympics and won the Asian Cup championship, it is only natural to think that
the Japanese soccer team has actually become stronger. However, the coach made
no dramatic changes to the lineup of the Japanese soccer team. So, what did Coach
Troussier tell his players?
Although we can guess the answer to some extent, let's imagine that you have been
named coach. What would you do to win more games? Some might prefer to concentrate
on devising a team formation strategy. Others might prefer to evaluate each player
and set individual goals for them, rather than concentrate on a formation strategy.
And still others might emphasize teaching skills such as effective through-passes.
All these answers have value, which makes it difficult to determine the best course
of action. Actually, it is not important to determine the best answer, but to
apply an appropriate answer for each situation in order to win games. More specifically,
all three strategies are indispensable to winning games: considering team formation
is a "collective level" strategy; setting goals for each player through individual
evaluations is an "individual level" strategy; and teaching effective skills is
a "knowledge level" strategy.
2. The Coach as System Designer
In the above discussion, we asked the reader to consider the role of soccer coach.
However, this paper is not intended to focus on soccer, but to address the design
of multiagent systems. What, then, is the relationship between soccer and multiagent
systems?
Considering that a soccer team comprises 11 players, the team can be considered
a multiagent system, with each player considered an autonomous agent. From this
perspective, the "coach" of the soccer team, as shown in Figure
1, corresponds to the "system designer" of a multiagent system. Just as a
good coach is expected to win, a "good system designer" is essential to the development
of a useful system.
3. What Should the System Designer Focus on?
What should the system designer focus on in order to develop a "good" system?
All system designers seek an answer to this very question. Unfortunately, currently
there is no consensus answer to this question. If one is developing a multiagent
system with many interacting agents, an answer to this question is urgently needed.
This is clear when one considers the difficulty of developing a useful multiagent
system Ñ even a slight change in the interaction between agents affects the multiagent
system as a whole. In spite of this difficulty, many companies still desire good
design methodologies for multiagent systems comprising several multiple systems,
such as telephone circuits or signals. This requirement is likely to increase
as systems become more complex.
Given this requirement, we focused on the importance of instructing soccer players
at the three levels to win games, and investigated their effectiveness in terms
of multiagent system design. More specifically, we investigated how the collective,
individual, and knowledge design levels influence multiagent systems. Through
an intensive computational simulation, we discovered the following implications:
Collective level: By analyzing the concept of organizational learning in
the context of organizational and management science, we discovered a factor that
affects multiagent systems at the collective level1.
Individual level: By investigating both the goals and evaluations of individual
agents, we discovered a factor that changes a characteristic of multiagent systems
at the individual level2.
Knowledge level: By exploring rules stored at the knowledge level, we discovered
a factor that resolves the trade-off between finding good solutions and reducing
computational costs3.
These implications play an important role at each level. This is because it is
difficult to determine a causal relationship among the factors that produce results
in multiagent environments. This indicates that the above three implications provide
a method for controlling the characteristics of micro-macro dynamics in which
interactions among agents at the micro level affect phenomena at the macro level,
and vice versa. However, this method is limited when only one of these implications
is employed, so we must consider all three levels in order to design good multiagent
systems4.
4. Can We Outdo Coach Troussier?
From the above discussion, we concluded that factors found at the three levels
have great potential for the development of good multiagent systems. Of course,
these factors alone cannot solve all problems, but they do contribute by supporting
design methodologies for developing useful multiagent systems. From these findings,
we believe that good multiagent systems can be developed with the factors we have
discovered, and we hope someday to propose good design methodologies that will
surpass Coach Troussier in effectiveness.
Finally, design methodologies for multiagent systems are also useful for understanding
and analyzing social and organizational phenomena resulting from complex interactions
among agents. For example, they allow us to explore answers to questions such
as, "What factors must be employed in order to improve the organizational performance?,"
and "How do we develop a robust company capable of coping with dynamic and complex
environments?" We addressed these issues from an organizational and management
viewpoint. We are also exploring ways of implementing autonomous agents in the
true sense by focusing on the self-generation of both goals and evaluation criteria,
a sense of values, and the boundary between self and others.
Please feel free to contact me if you are interested in our work or if you have
any questions.
Reference
