If multi-agent systems are to become widely accepted as a basis for large scale applications, adequate agent concepts will be essential. To address this issue, we show in this paper that agents need to have communication concepts and organization concepts. We argue that instead of the usual approach of starting from a set of intentional states, the intentional structure should be deduced in terms of interaction. To this end, we come up with ontologies related to communication and organization. Unlike most previous work, the presented study shows a new way to combine agent and ontology technologies. Indeed, ontologies have established themselves as a powerful tool to enable knowledge sharing, and a growing number of applications have benefited from the use of ontologies as a means to achieve semantic interoperability among heterogeneous, distributed systems. Ontologies are used in this paper as natural arguments to belief-goal-role agents. We illustrate the concepts by the well known prey/predator example.
Artificial intelligence (AI) has emphasized building "stand-alone
systems" that can solve problems with minimal help from other
systems (computer or human). These systems have traditionally
been brittle, in the sense that they fail miserably when presented
with problems even slightly outside of their range of expertise[
The purpose of this paper is to establish how these different
models stand in relation to each other and contemporary concepts
of agency, and in particular it presents a new approach which is
different from most known in AI field: while the
Belief-DesireIntention architecture draws its inspiration from the philosophical
theories of Bratman [
Actually, the conception of multi-agent systems covers many meanings each referring to a peculiar trend in research. These trends can be grouped in two standpoints: individual1 conception and mass conception.
1 This term was used by Ferber in [
This point of view gathers all researchers who think that the
solution goes only through the formal representation of an agent
model [
Unlike individual conception, this approach does not mask the
original aspect of multi-agent systems which is cooperation.
Cooperation is often presented as one of the key concepts which
differentiates multi-agent systems from other related disciplines
such as distributed computing, object-oriented systems, and expert
systems [
Our approach is in keeping with the latter standpoint. Indeed,
starting from the study of cooperation in multi-agent systems [
We support the idea that communication as well as organization are mainstays of cooperation.
In this paper, we use a two steps approach: - Conceptualization which consists in focusing our attention on the system structures that seems relevant to the problem to be solved. Then, describing informally those structures in a language called conceptualization language. - formal specification which consists in the formalization of the conceptualization. This definition is consistent with the usage of ontology .
This approach is applied for communication as well as organization.
This paper is organised as follows. In the next section, we present an ontology for communication. Section 3 provides an organization ontology. Section 4 identifies the agent concepts. We illustrate our work with the well known prey/predator problem.
Although communication is not universally assumed in
multiagent systems research, we argue that it is fundamental to the
everyday process of cooperation. Communication in multi-agent
systems as for human beings, is the basis of interaction and social
organization. Communication consists of a set of psychological
and physical processes relating the sender to the addressee(s) in
order to reach some goals [
The goal of this section is to derive the underlying concepts to a
communication model appropriate to multi-agent systems.
Let C be a communication model. C is based on speech act
theory[
In more traditional AI terms, an agent's beliefs are its knowledge. Beliefs change due to the external environment of the agent, his communication with the others and his proper reasoning. Beliefs are acquired through perception, communication, deduction, and prediction.
The concept of belief is an essential element of C. Indeed, beliefs can be a subset of the propositions that agents exchange. Generally, communication depends on the beliefs and the application domain.
Example 1. - In the prey/predator problem, the prey's position on the grid is a predator’s belief. Whenever the prey gets into the perception field of a predator (hence he believes knowing its position), he communicates this information to the other predators. - A communication can be due to the fact that a predator believes not knowing the prey's position. In this case, he sends a request to his colleagues asking for the prey's position. 2 We adopt this definition because we think that it is appropriate to multi-agent systems problematics.
A goal is an underlying concept to the psychological level of a communication. It refers to a potential state of the world to which the agent belongs. The achievement of a goal is at the beginning of an agent behaviour (including his communicative actions). A goal may be local or global and can be expressed in terms of the agent beliefs.
For a predator, getting closer to a prey from the southern side is a local goal which helps reaching the global goal consisting in the capture of the prey. This local goal can be expressed in terms of the following beliefs: the prey's position and the predator's position on the grid.
As previously stated, a communication is defined as a speech act and a-fortiori as an action. Accordingly, an action is the basic element of a communication model. Actions are interactions between an agent and the external world. An agent can perform at a point of time one of the following actions: - physical actions are interactions between agents and the spatial environment, - communicative actions are interactions between agents. They can be emission or reception actions, - private actions are internal functions of an agent. They correspond to an agent exploiting its internal computational resources, and - decision action can generate communicative, physical and private actions. A decision action can also update the agent's beliefs. We assume that the agent's goals are modified only after a negotiation with the other agents.
The actions to execute are determined by the resolution methods and communication protocols.
We denote by APH the set of physical actions, APR the set of private actions, ACO the set of communicative actions such that ACO = ACOE ∪ ACOR where ACOE is the set of emission actions and ACOR is the set of reception actions, τ a decision action, which an agent can execute.
- In the prey/predator game, moving on the grid is a physical
action.
- informing another predator is a communicative action.
Concerning the actions execution, an agent use an interaction
paradigm (e.g. an agenda). The interaction paradigm implicitly
defines the metaphors used by the agents to interact and cooperate
among each other [
In the definition of communication, three dimensions have been distinguished among which the psychological dimension (whose concepts are beliefs, goals, and actions) and the physical dimension (which support communication). The transition step between the afore mentioned dimensions is message production. In a multi-agent universe, a message is a specification of a speech act to which we can add the physical processes of communication which support message transmission. Practically, a message can be considered as a four attributes structure: elocution force, propositional content, sender, and routing. Figure 1 represent an example of this structure according to the prey/predator problem.
Whenever we have to deal with task distribution and well-knit
interaction between agents in a multi-agent system, the basic
problem is an organizational issue i.e. deciding who will do what
and when [
Organization in animal or human societies deals with task distribution and work division according to the proper skills of each agent member. Next, we draw the key concepts related to organization. An organizational model defines how to dispatch the tasks between cooperative agents and their possible relations.
A problem is defined by a set of tasks. Each task can be optional or compulsory according to the resolution strategy (see § 3.4). A task is an ordered set of actions which allows to reach a goal.
Definition 3.1 [A problem] Let T={t1, t2, …, ti, …, tn} be a set of tasks defining the problem to be solved,
TC: a restriction of T to compulsory tasks, and
TO: a restriction of T to optional tasks, and
T = T C ∪ TO and TC ∩ TO = ∅ .
The distinction between compulsory tasks and optional ones can be justified by the fact that it is a better way to simulate the reality.
The prey/predator problem includes the following compulsory tasks: "follow the prey", "communicate the prey’s position once it is seen". To this set, we can add the following optional tasks: "communicate the predator's position at each step", or "communicate the distance covered by the predator".
3.2 Role In a cooperation framework, it is necessary to define the respective roles that the agents play in a resolution process. The role characterizes the responsibilities which would be attributed to an agent i.e., the tasks he is charged with. An agent's role is defined by a subset of compulsory tasks and a subset of optional tasks.
Definition 3.2 [An agent role] Let a be an agent, TCa be the compulsory tasks of a, TOa be the optional tasks of a. The role of the agent a is defined by ra such that: ra = < TCa, T Oa >.
The concept of cooperation mode has been previously stated in
[
In the prey/predator game, once the prey's position is known by all the predators, the cooperation mode related to this situation is: <"surround the prey from the south", "surround the prey from the north", "surround the prey from the east", "surround the prey from the west">.
The relevance of a cooperation mode is measured by the cooperation strategy which will be described in the following paragraph.
In a distributed problem solving, each agent member of the system is able to process a set of tasks by applying some resolution methods within his competences.
Agent competences can be complementary or redundant. For the same tasks, an agent may be more competent than another (he puts into practice a resolution method more efficaciously). For other tasks, the agents can have different methods but there is no way to determine a priori which method is the most efficient. Accordingly, to select an agent to achieve a task, the resolution context should be taken into account. Hence, a dynamic distribution of tasks seems to be more appropriate than a static distribution. If the abilities of the agents are redundant, then they can play the same role. In this case, an agent should be selected to accomplish each role. According to the context, we can specify different indications concerning the way the choice is done: for example, in the case of system-human cooperation, we can indicate that the user must play all the decision roles even if the system is competent to accomplish such roles.
We call a cooperation strategy, a process which determines the appropriate cooperation mode to a given situation. In this case, we talk about dynamic organization3.
In the prey/predator game, a cooperation mode where an agent’s role is "getting closer to a prey from the north" can be substituted for a cooperation mode where the same agent play the following role: "getting closer to a prey from the east". It is because of the respective positions of the predators on the grid, that the cooperation strategy administer this transition in cooperation mode.
In order to guarantee the coherence of an organizational model,
some properties should be verified. We don’t address those
properties in this paper, but we think that they are closely related
to the application domain.
3 Mike Wooldridge & al., support the idea of dynamic
organization when they talk about reactive or on-fly cooperation
[
Cooperation is perhaps the paradigm example of social activity in
both real and artificial social systems; it is certainly the best
studied process in multi-agent systems research [
A conceptualization of cooperation has been presented so far. Indeed, two aspects have been mentioned as mainstays of cooperation: communication and organization. Concerning communication, the terminology includes the concepts of beliefs, goals, actions and a message. Organization is dynamic because of the concept of cooperation strategy.
Obviously, an organizational structure in which roles distribution is dynamic increase communication cost, but can give better results for some complex problems. A static structure doesn't cope with the environment hazards and limits the possible configurations that can occur in the resolution process4. Multi-agent systems put into practice a set of techniques and concepts allowing some concrete (e.g. robots) or abstract (software) entities called "agents" to cooperate according to some cooperation modes. By focusing on interaction and individual satisfaction, multi-agent systems ban thinking in a centralized or global way5. Hence, we keep from our previous study the following concepts for an agent: beliefs and goals as communication concepts and roles as concepts related to organization.
The agent metaphor comes packaged with a number of powerful psychological abstractions such as Beliefs, Desires, and Intentions. The work presented explores a particular type of rational agent: a Belief-Goal-Role agent. This making up of the intentional structure of an agent was subject to argumentation in this paper. The development of an ontology for cooperation has ended in identifying beliefs, goals as communication concepts and roles as organization concepts. Unlike most previous work, the presented study shows a new way to combine agent and ontology technologies. Indeed, ontologies have established themselves as a powerful tool to enable knowledge sharing, and a growing number of applications have benefited from the use of ontologies as a means to achieve semantic interoperability among heterogeneous, distributed systems. Ontologies are used in this paper as natural arguments to belief-goal-role agents.
Finally, we think that the study presented in this paper open new perspectives on the realization of agent systems. The ontology developed for communication as well as for organization draws its inspiration from the widely recognized fact that interaction is the most important single characteristic of complex systems. Many researchers believe that in future, computation itself will be understood chiefly as a process of interaction. 4 In certain kinds of problems, a static structure may be more appropriate than a dynamic one. Such discussion is beyond the scope of this paper; it can be addressed apart in another study with more details. Generally, and according to the essence of multiagent systems, a dynamic structure can respond better to complexity. 5 Accordingly, it is forbidden to talk about cooperation mode (see section 3) whenever we talk about agent systems.