AN ATTITUDE BASED MODELING OF
AGENTS IN COALITION
Madhu Goyal
Faculty of Information Technology
University of Technology Sydney, PO BOX 123, Broadway, NSW 2007, Australia
Keywords: Multi-agent, Coalition formation, Attitudes.
Abstract: One of the main underpinning of the multi-agent systems community is how and why autonomous agents
should cooperate with one another. Several formal and computational models of cooperative work or
coalition are currently developed and used within multi-agent systems research. The coalition facilitates the
achievement of cooperation among different agents. In this paper, a mental construct called attitude is
proposed and its significance in coalition formation in a dynamic fire world is discussed. This paper
presents ABCAS (Attitude Based Coalition Agent System) that shows coalitions in multi-agent systems are
an effective way of dealing with the complexity of fire world. It shows that coalitions explore the attitudes
and behaviors that help agents to achieve goals that cannot be achieved alone or to maximize net group
utility.
1 INTRODUCTION
Coalition formation is an important cooperation
method in multi-agent systems. A coalition, is a
group of agents who join together to accomplish
a task that requires joint task execution which
otherwise be unable to perform or will perform
poorly. It is becoming increasingly important as it
increases the ability of agents to execute tasks and
maximize their payoffs. Thus the automation of
coalition formation will not only save considerable
labour time, but also may be more effective at
finding beneficial coalitions than human in complex
settings. To allow agents to form coalitions, one
should devise a coalition formation mechanism that
includes a protocol as well as strategies to be
implemented by the agents given the protocol.
This paper will focus on the issues of coalitions in
dynamic multi-agent systems: specifically, on issues
surrounding the formation of coalitions among
possibly among heterogeneous group of agents, and
on how coalitions adapt to change in dynamic
settings. Traditionally, an agent with complete
information can rationalize to form optimal
coalitions with its neighbors for problem solving.
However, in a noisy and dynamic environment
where events occur rapidly, information cannot be
relayed among the agent frequently enough,
centralized updates and polling are expensive, and
the supporting infrastructure may partially fail,
agents will be forced to form sub-optimal coalitions.
Similarly, in such environments, changes in
environmental dynamics may invalidate some of the
reasons for the original existence of a coalition. In
this case, individual agents may influence the
objectives of coalition, encourage new members and
reject others and the coalition as a whole adapts as a
larger organism. In such settings, agents need to
reason, with the primary objective of forming a
successful coalition rather than an optimal one, and
in influencing the coalition (or forming new
coalitions) to suit its changing needs. This includes
reasoning about task allocation, the needs of self and
others, information exchange, uncertainty and
information incompleteness, coalition formation
strategies, learning of better formation strategies,
and others.
Coalition formation has been addressed in game
theory for some time. However, game theoretic
approaches are typically centralized and
computationally infeasible. MAS researchers (Kraus
et al 2003) (Sandholm et al, 1999) (Shehory and
Kraus, 1995) (Li et al, 2003), using game theory
concepts, have developed algorithms for coalition
formation in MAS environments. However, many of
165
Goyal M. (2007).
AN ATTITUDE BASED MODELING OF AGENTS IN COALITION.
In Proceedings of the Ninth International Conference on Enterprise Information Systems - AIDSS, pages 165-170
DOI: 10.5220/0002361401650170
Copyright
c
 SciTePress

them suffer from a number of important drawbacks
like they are only applicable for small number of
agents and not applicable to real world domains.
This paper introduces ABCAS, a novel attitude
based coalition agent system in the fire world. The
task of fire fighting operations in a highly dynamic
and hostile environment is a challenging problem.
We suggest a knowledge-based approach to the
coalition formation problem for fire fighting
missions. Thus the objective of this paper is to
design and develop an attitude based approach to the
coalition formation for fire fighting problem that
would help them to accomplish their tasks during the
fire. Owing to the special nature of this domain,
developing a protocol that enables agents to
negotiate and form coalitions, and provide them with
simple heuristics for choosing coalition partners is
quite challenging task. The protocol allows the
agents to form coalitions, and provide them with
simple heuristics that allow the agents to form
coalitions in face of time constraints and incomplete
information.
2 A FIRE WORLD
We have implemented our formalization on a
simulation of fire world FFWorld (Goyal, 2004)
using a virtual research campus. FFWorld is a
dynamic, distributed, interactive, simulated fire
environment where agents are working together to
solve problems, for example, rescuing victims and
extinguishing fire. In a world such as this, no agent
can have full knowledge of the whole world.
Humans and animals in the fire world are modeled
as autonomous and heterogeneous agents. While the
animals run away from fire instinctively, the fire
fighters can tackle and extinguish fire and the
victims escape from fire in an intelligent fashion. An
agent responds to fire at different levels. At the
lower level, the agent burns like any object, such as
chair. At the higher level, the agent reacts to fire by
quickly performing actions, generating goals and
achieving goals through plan execution.
This world contains all the significant features of
a dynamic environment and thus serves as a suitable
domain for collaborating agents. Agents in the fire
domain do not face the real time constraints as in
other domains, where certain tasks have to be
finished within the certain time. However, because
of the hostile nature of the fire, there is strong
motivation for an agent to complete a given goal as
soon as possible. There are three main objectives for
intelligent agents in the world during the event of
fire: self-survival, saving objects including lives of
animals and other agents and put-off fire. Because of
the hostile settings of the domain, there exist a lot of
challenging situations where agents need to do the
cooperative activities. Whenever there is fire, there
is need of coalition between the fire fighters (FF-
agent), volunteers (Vol-agent) and victim agents
(Vic-agent)(Fig. 1). The fire fighters perform all the
tasks necessary to control an emergency scene. The
problem solving activities of the fire fighters are
putting out fire, rescuing victims and saving
property. Apart from these primary activities there
are a number of sub tasks eg. run towards the exit,
move the objects out of the room, remove obstacles,
and to prevent the spread of fire. The first and
paramount objective of the victim agents is self-
survival. The role of volunteer agents is to try to
save objects from the fire and help out other victims
who need assistance when they believe their lives
are not under threat. To achieve these tasks there is
need of coalitions between these agents is necessary.
Figure 1: Coalition between Fire-fighter, Volunteer and
Victim Agent.
3 STRATEGIC COALITIONS IN
AN AGENT BASED HOSTILE
WORLD
The coalition facilitates the achievement of
cooperation among different agents. The cooperation
among agents succeeds only when participating
agents are enthusiastically unified in pursuit of a
common objective rather than individual agendas.
We claim that cooperation among agents is achieved
only if the agents have a collective attitude towards
cooperative goal as well as towards cooperative
plan. From collective attitudes, agents derive
individual attitudes that are then used to guide their
behaviours to achieve the coalition activity. The
agents in a coalition can have different attitudes
ICEIS 2007 - International Conference on Enterprise Information Systems
166

depending upon the type of the environment the
agent occupies.
3.1 Definition of Attitude
Attitude is a learned predisposition to respond in a
consistently favourable or unfavourable manner
with respect to a given object (Fishbein and
Ajzen,1975
). In other words, the attitude is a preparation in
advance of the actual response, constitutes an
important determinant of the ensuing behaviour.
However this definition seems too abstract for
computational purposes. In AI, the fundamental
notions to generate the desirable behaviours of the
agents often include goals, beliefs, intentions, and
commitments. Goal is a subset of states, and belief is
a proposition that is held as true by an agent.
Bratman (Bratman, 1987) addresses the problem of
defining the nature of intentions. Crucial to his
argument is the subtle distinction between doing
something intentionally and intending to do
something. The former case might be phrased as
deliberately doing an action, while intending to do
something means one may not be performing the
action in order to achieve it. Cohen and Levesque
(Cohen and Levesque, 1991), on the other hand,
developed a logic in which intention is defined.
They define the notion of individual commitment as
persistent goal, and an intention is defined to be a
commitment to act in a certain mental state of
believing throughout what he is doing. Thus to
provide a definition of attitude that is concrete
enough for computational purposes, we model
attitude using goals, beliefs, intentions and
commitments. From the Fishbein’s definition
(Fishbein and Ajzen,1975) it is clear that when an
attitude is adopted, an agent has to exhibit an
appropriate behaviour (predisposition means behave
in a particular way). The exhibited behaviour is
based on a number of factors. The most important
factor is goal or several goals associated with the
object. During problem solving, an agent in order to
exhibit behaviour may have to select from one or
several goals depending on the nature of the
dynamic world.
In a dynamic multiagent world, the behaviour is
also based on appropriate commitment of the agent
to all unexpected situations in the world including
state changes, failures, and other agents’ mental and
physical behaviours. An agent intending to achieve a
goal must first commit itself to the goal by assigning
the necessary resources, and then carry out the
commitment when the
appropriate opportune comes.
Second, if the agent is committed to executing its
action, it needs to know how weak or strong the
commitment is. If the commitment is week, the
agent may not want to expend too much of its
resources in achieving the execution. The agent thus
needs to know the degree of its commitment towards
the action. This degree of commitment quantifies the
agent’s attitude towards the action execution. For
example, if the agent considers the action execution
to be higher importance (an attitude towards the
action), then it may choose to execute the action
with greater degree of commitment; otherwise, the
agent may drop the action even when it had failed at
the first time. Thus, in our formulation, an agent
when it performs an activity, since the activity is
more likely that it will not succeed in a dynamic
world; agents will adopt a definite attitude towards
every activity while performing that activity. The
adopted attitude will guide the agent in responding
to failure situations. Also the behaviour must be
consistent over the period of time during which the
agent is holding the attitude. Thus attitudes, once
adopted, must persist for a reasonable period of time
so that other agents can use it to predict the
behaviour of the agent under consideration. An
agent cannot thus afford to change its attitude
towards a given object too often, because if it does,
its behaviour will become somewhat like a reactive
agent, and its attitude may not be useful to other
agents. Once an agent chose to adopt an attitude, it
strives to maintain this attitude, until it reaches a
situation where the agent may choose to drop its
current attitude towards the object and adopt a new
attitude towards the same object. Thus we define
attitude as: An agent’s attitude towards an object is
its persistent degree of commitment to one or several
goals associated with the object, which give rise to
persistent favourable or unfavourable behaviour to
do some physical or mental actions.
3.2 Type of Attitudes
The attitudes of the agents in the world consist of
attitudes towards the physical objects, mental objects,
processes and other agents. When attitudes are
attached to physical objects, the agents are able to
evaluate the liking, importance or location etc. of
these physical objects. When attitudes are attached
to mental objects, agents are able to communicate
and reason with those mental objects. For example,
agents can actively monitor their plans so those
plans can be re-organised or abandoned when the
world state changes. If the object denotes a mental
object such as a plan, higher-priority can be an
AN ATTITUDE BASED MODELING OF AGENTS IN COALITION
167

attitude that the agent may hold towards the plan. In
that case, the agent will perform behaviour
appropriate to this attitude, which may involve
physical, communicative, and mental actions or a
combination of these which may lead to behaviour
where the agent gives higher preference to the plan
compared to the other plans in all possible situations.
Agents can also have attitudes towards processes
such as execution of actions and plans, the process
of achieving goals, etc. For example, if the execution
of a plan goes on for too long, appropriate attitude is
necessary to define how to handle the situation.
Behaviours exhibited by an agent in a
multiagent environment can be either individualistic
or collective. Accordingly, we can divide attitudes in
two broad categories: individual attitudes and
collective attitudes. The individual attitudes
contribute towards the single agent’s view towards
an object or person. An agent’s attitude toward an
object is based on its salient beliefs about that object.
The agent’s individual attitude toward a fire world,
for example, is a function of its beliefs about the fire
world. The collective attitudes are those attitudes,
which are held by multiple agents. The collective
attitudes are individual attitudes so strongly
interconditioned by collective contact that they
become highly standardised and uniform within the
group, team or society etc. The agents can
collectively exist as societies, groups, teams, friends,
foes, or just as strangers, and collective attitudes are
possible in any one of these classifications. For
example, the agents in the collection called friends,
can all have a collective attitude called friends,
which is mutually believed by all agents in the
collection. A collective attitude can be viewed as an
abstract attitude consisting of several component
attitudes, and for an individual agent to perform an
appropriate behaviour; it must hold its own attitude
towards the collective attitude. Thus, for example, if
A1 and A2 are friends, then they mutually believe
they are friends, but also each Ai must have an
attitude towards this infinite nesting of beliefs so
that it can exhibit a corresponding behaviour. Thus,
from A1’s viewpoint, friends is an attitude that it is
holding towards the collection {A1, A2} and can be
denoted as friendsA1(A1, A2). Similarly, from A2’s
view point, its attitude can be denoted as
friendsA2(A1,A2).
3.3 Attitude Based Agents
We adopt a BDA (Fig.2 modified BDI) based
approach in which agent is comprised of: beliefs
about itself, others and the environment; set of
desires representing the states it wants to achieve;
and attitudes corresponding to the plans adopted in
pursuit of the desires. In comparison to traditional
BDI (Cohen and Levesque,1991) model, we have
replaced intentions with attitudes. We say that
intentions are primitive forms of attitudes without
degree of commitment in them. An agent has a set of
attitudes, each with a degree of commitment which
persists according to the current situation. The
attitudes are represented by following attributes:
Figure 2: BDA Agent Architecture.
Name of Attitude: This attribute describes the name
of the attitude e.g. like, hate, cautious etc.
Description of Object: The description of the object
contains the name of the object and a description of
the internal organization in terms of the components
of the object.
Basic agent behaviour towards x: This attribute
specifies the behaviour that will be performed by the
agent with respect to the object x.
Evaluation: This attribute specifies whether the
attitude is favourable or not.
Concurrent attitudes: This attribute specifies any
other attitudes that can coexist with this attitude.
Persistence of Attitude: This attribute specifies how
long the attitude will persist under various situations.
For example, it may specify how the attitude itself
will change over time; that is, when to drop it and
change it to another attitude, when to pick it up and
how long to maintain it.
Type of Attitude: This attribute specifies whether the
attitude is individual or collective.
3.4 Attitude based Coalition Agent
Model
We claim that successful coalition is achieved only
if the agents have coalition as a collective abstract
attitude. From this collective attitude, agents derive
ICEIS 2007 - International Conference on Enterprise Information Systems
168

individual attitudes that are then used to guide their
behaviors to achieve the coalition. Suppose there n
agents in a coalition i.e. A
1
…. A
n
. So the collective
attitude of the agent A
1
…. A
n
towards the coalition
is represented as Coal
A1..An
(A
1
, .. ,A
n
). But from A
1
’s
viewpoint, team is an attitude that it is holding
towards the collection (A
1,
…,A
n
) and can be denoted
as Coal
A1
(A
1
, A
2
). Similarly from A
n
’s viewpoint, its
attitude can be denoted as Coal
An
(A
1
,.., A
n
). But the
collective attitude Coal
A1 An
(A
1
,…, A
n
) is
decomposed into the individual attitudes only when
all the agents mutually believe that they are in the
coalition. The coalition attitude can be represented
in the form of individual attitudes towards the
various attributes of the coalition i.e. coalition
methods, coalition rule base, and coalition
responsibility.
The attitudes of an agent existing in a coalition
consist of attitude towards coalition as well as
attitude towards coalition activity. At any time, an
agent may be engaged in one of the basic coalition
activities i.e. coalition formation, coalition
maintenance, and coalition dissolution. Instead of
modelling these basic activities as tasks to be
achieved, we have chosen to model them as attitudes.
Coalition (A1,..,A2)
This attitude is invoked when the agents are in a
team state. This attitude guides the agents to perform
the appropriate coalition behaviours.
Name of Attitude: Coalition
Description of Object: (1) Name of Object: set of
agents (2) Model of Object: {A1,An | Ai is an agent}
Basic agent behaviour: coalition behaviour specified
by agent’s rule base
Evaluation: favourable
Persistence: This attitude persists as long as the
agents are able to maintain it.
Concurrent attitudes: all attitudes towards physical
and mental objects in the domain.
Type of Attitude: collective.
3.4.1 Coalition Formation
In the fire world, the event triggering the coalition
formation process is a fire. Whenever there is fire,
the security officers call the fire-fighting company to
put out the fire. Then the fire fighters arrive at the
scene of fire and get the information about when,
how and where the fire had started. Suppose there is
a medium fire in the campus, which results in the
attitudes medium-fire and dangerous-fire towards
the object fire. The attitude Coal-form is also
generated, which initiates the team formation
process. We propose a dynamic team formation
model, in which we consider initially the mental
state i.e. the beliefs of all the agents is same. The
fire-fighting agents recognise appropriateness of the
team model for the task at hand; set up the
requirements in terms of other fellow agents, role
designation, and structure; and develop attitudes
towards the team as well as towards the domain.
In order to select a member of the team, our agent
will select the fellow agent who has following
capabilities:
- Has knowledge about the state of other agents.
- Has attitude towards the coalition formation.
- Can derive roles for other agents based on skills
and capabilities.
- Can derive a complete joint plan.
- Can maintain a coalition state.
Our method of forming a coalition is like this; the
agents start broadcasting message to other agents
“Let us form a coalition”. The agents will form a
coalition if two or more than two agents agree by
saying, “Yes”. If the agent do not receive the “Yes”
message, it will again iterate through the same steps
until the coalition is formed. The coal-form is
maintained as long as the agents are forming the
team. Once the team is formed, agents will drop the
coal-form attitude and form the coal attitude,
which
will guide the agents to produce various team
behaviours.
Coal-form (A
1
, ..,A
2
)
This attitude is invoked when the agents have to
form a coalition to solve a complex problem.
Name of Attitude: Coal-form
Description of Object: (1) Name of Object: set of
agents. (2) Model of Object: {A1,An | Ai is an agent}
Basic agent behaviour: invokes coalition formation
rules.
Evaluation: favourable
Persistence: The agent holds this attitude as long as
it believes that a coalition formation is possible.
Concurrent attitudes: All attitudes towards physical
and mental objects in the domain.
Type of Attitude: individual
3.4.2 Coalition Maintenance and Dissolution
While solving a problem (during fire fighting
activity) the coalition agents have also to maintain
the coalition. During the coalition activity the agents
implement the coalition plan to achieve the desired
coalition action and sustain the desired
AN ATTITUDE BASED MODELING OF AGENTS IN COALITION
169

consequences. The coaltion maintenance behaviour
requires what the agent should do so that coalition
does not disintegrate. In order to maintain the
coalition each agent should ask the other agent
periodically or whenever there is a change in the
world state, whether he is in the coalition. So the
attitudes like periodic-coalition-maintenance and
situation-coalition-maintenance are produced
periodically or whenever there is a change in the
situation. These attitudes help the agent to exhibit
the maintenance behaviours.
When the team task is achieved or team activity
has to be stopped due to unavoidable circumstances,
the attitude coal-unform is generated. This attitude
results in the dissolution of the team and further
generates attitude escape. For example, when the
fire becomes very large, the agents have to abandon
the team activity and escape. The attitude coal–
unform is maintained as long as the agents are
escaping to a safe place. Once the agents are in the
safe place, the attitudes team-unform and escape are
relinquished. In case the fire comes under control,
the agents again form a team by going through the
steps of team formation.
4 CONCLUSIONS
This paper has developed a novel framework for
managing coalitions in a hostile dynamic world.
Coalition is guided by the agent’s dynamic
assessment of agent’s attitudes given the current
scenario conditions, with the aim of facilitating the
agents in coalitions to complete their tasks as
quickly as possible. In particular, it is outlined in this
paper that how agents can form and maintain a
coalition, and how it can offers certain benefits to
cooperation. Our solution provides a means of
maximizing the utility and predictability of the
agents as a whole. Its richness presents numerous
possibilities for studying different patterns of
collaborative behaviour.
REFERENCES
Bratman.M.E.,1987 Intentions, Plans and Practical Reason.
Harvard University Press, Cambridge, MA.
Cohen. P. R. and Levesque. H. J., 1991. Teamwork.
Special Issue on Cognitive Science and Artificial
Intelligence, 25(4).
Fishbein. M and Ajzen. I, 1975. Belief, Attitude, Intention
and Behaviour: An Introduction to theory and
research. Reading, MA,USA: Addison-Wesley.
Goyal. M,2004. Collaborative Negotiations in a Hostile
Dynamic World. The International Conference on
Artificial Intelligence (IC-AI’ 04), Las Vegas, USA.
Kraus.S, O. Shehory, and G. Taase,2003. Coalition
formation with uncertain heterogeneous information,
in Proceedings of the 2nd International Conference on
Autonomous Agents and Multi-agent Systems (AAMAS
2003), Melbourne, Australia ,July 2003.
Sandholm, T., Larson, K., Andersson, M., Shehory, O.,
and Tohmé, F.,1999. Coalition Structure Generation
with Worst Case Guarantees. Artificial Intelligence,
111(1-2), 209-238.
Shehory. O, Sarit Kraus,1995. Task Allocation via
Coalition Formation Among Autonomous Agents.
Proceedings of the Fourteenth International Joint
Conference on Artificial Intelligence (IJCAI-95).
Li. C, S. Chawla, U. Rajan, and K. Sycara, 2003.
Mechanisms for Coalition Formation and Cost Sharing
in an Electronic Marketplace. Tech. report CMU-RI-
TR-03-10, Robotics Institute, Carnegie Mellon
University.
ICEIS 2007 - International Conference on Enterprise Information Systems
170