Enhancing Believability of Virtual Soccer Players: Application of a BDI-Model with Emotions and Trust - Developing Concepts in Applied Intelligence - page 124

In addition to the scenario shown above, various other scenarios have been
simulated in a systematic manner (varying from scenarios in which only positive or
negative events occur to scenarios in which both positive and negative events occur).
Due to space limitations, the details of these simulations are not shown here. To
improve the model's accuracy, each simulation run has been evaluated in the
following manner. Based on the literature in decision making, emotion, and trust (see,
e.g., [4, 5] and references in those papers), a number of requirements for the behavior
of the agents have been formulated; some examples are the following:

• 'events that contribute to an agent's goals lead to an increase of happiness'

• 'agents that perform more successful actions are trusted more'

'happy agents perform different actions than sad agents'
These requirements have been tested against all simulated traces. In case a
requirement was not fulfilled, small modifications in the model were made, either by
correcting bugs in the specification or (more often) by adapting the values of the
parameters involved. After a number of iterations, the model was considered ready to
be implemented in a more realistic setting.

•

4 RoboCup Application

After the initial tests mentioned above pointed out that the model showed acceptable
behavior, it has been implemented within the RoboCup 2D virtual soccer
environment. RoboCup is an international competition founded in 1997 with the aim
to develop autonomous (hardware and software) agents for various domains, with the
intention to promote research and education in Artificial Intelligence [10]. RoboCup
has four competition domains (Soccer, Rescue, Home, and Junior), each with a
number of leagues. For the current project, the RoboCup Soccer 2D Simulation
League was used, since this league provides good possibilities for rapid prototyping
for games with many inter-player interactions.
In order to implement the model for decision making with emotions and trust in an
efficient manner, part of the existing code of the 2005 Brainstormers team [14] was
reused. This code contains a number of explicit classes for different actions (such as
'shoot_at_goal', 'pass_from_to', 'dribble', 'tackle', and 'run_free'). To create optimal
strategies at the level of the team, the Brainstormers used Machine Learning
techniques to learn which actions were most appropriate in which situations. Instead,
for the current paper, a different approach was taken: for each individual agent, the
original action classes were directly connected to the output (the generated actions) of
the implemented decision making model. Thus, instead of having a team of agents
that act according to a learned collective strategy, each agent in our implementation
acts according to its personal role, (partly emotion-driven) desires and intentions 1 .
This approach turned out to be successful in generating (intuitively) realistic soccer
matches, in which the behavior of the players is influenced by their emotions and trust
states. A screenshot of the application is shown in Figure 4. The top of the figure

1 Recall that the aim of the current paper is not to develop optimal strategies, but rather to
enhance believability of the virtual players.