EXTENDING MAS-ML TO MODEL PROACTIVE

AND REACTIVE SOTWARE AGENTS

Enyo José Tavares Gonçalves

Instituto Federal de Educação Ciência e Tecnologia do Ceará, Maracanaú, CE, Brazil

Mariela I. Cortés, Gustavo A. L. de Campos

Universidade Estadual do Ceará, Fortaleza, CE, Brazil

Viviane Torres da Silva

Universdade Federal Fluminense, Niterói, RJ, Brazil

Keywords: Proactive and Reactive Agents, Multi-agent System, Modelling Language, Conceptual Framework.

Abstract: The existence of Multi Agent System (MAS) where agents with different internal architectures interact to

achieve their goals promotes the need for a language capable of modeling these applications. In this context

we highlight MAS-ML, a MAS modeling language that performs a conservative extension of UML while

incorporating agent-related concepts. Nevertheless MAS-ML was developed to support pro-active agents.

This paper aims to extend MAS-ML to support the modelling of not only proactive but also reactive agents

based on the architectures described in the literature.

1 INTRODUCTION

Nowadays, the agent technology has been widely

applied to solve a vast set of problems. Russell and

Norvig (2003) define an agent as an entity that can

perceive its environment through sensors and act in

environment through actuators. Unlike objects, agents

are more complex entities with behavioural properties,

such as: (i) they are autonomous and not passive, and

(ii) able to interact through exchange of messages and

not by explicit task invocation (Wagner, 2003). Multi-

Agent System (MAS) is the sub-area of Artificial

Intelligence that investigates the behaviour of a set of

autonomous agents, aiming to resolve a problem that is

beyond the capacity of a single agent (Jennings, 1996).

The agent-oriented development paradigm requires

adequate techniques to explore its benefits and

features, in order to support the construction and

maintenance of this type of software (Zambonelli et al.,

2001). A simple agent is classified according to its

internal architecture that determines distinct agency

properties, attributes and mental components. These

features introduce additional complexity to the system

development. A MAS may encompass multiple types

of agents with different internal architectures (Weiss,

1999). Thus, the existence of a language to support the

modelling of different internal agent architectures is

strongly desirable

Several modelling languages have been proposed

in the literature to model agents and their systems. One

of them is called MAS-ML (Multi-Agent System

Modelling Language) (Silva and Lucena, 2004) (Silva,

Choren and Lucena, 2008a) that performs a

conservative extension to UML based on the agent-

oriented concepts defined in the conceptual framework

TAO (Taming Agents and Objects) (Silva and Lucena,

2004). In particular, the following characteristics of the

language can be highlighted: (i) the support for the

modelling of main MAS entities: agents, organization

and environments; (ii) the support for conventional

objects; (iii) the support for modelling static and

dynamic properties; (iv) the modelling of agent roles,

that are important while defining agent societies; and

(v) the clear justified extension of the UML

metamodel to model agent-related properties based on

TAO (Silva, Choren and Lucena, 2008a). Due to its

characteristics, MAS-ML is known as one of the main

adequate modelling language to model MAS.

José Tavares Gonçalves E., I. Cortés M., A. L. de Campos G. and Torres da Silva V. (2010).

EXTENDING MAS-ML TO MODEL PROACTIVE AND REACTIVE SOTWARE AGENTS.

In Proceedings of the 12th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

75-84

DOI: 10.5220/0002907700750084

 SciTePress

MAS-ML was originally designed to support the

modelling of only proactive agents that are goal-

oriented entities and guided by pre-established plans.

However, not all MAS require or permit their agents

are pro-active, as the case of simulations for an ant

colony (Dorigo and Stützle, 2004). Create goal-based

agents with plan in stochastic and partially observable

environments can be a very complex task (Weiss,

1999). Therefore, it is fundamental to extend MAS-

ML to be able to model not only proactive agents that

have pre-defined plans but also reactive ones. In

addition, MAS-ML should also be able to model

proactive agents able to create new plans and that use

utility functions to execute (Russell and Norvig, 2003).

In this paper, we describe an extension of the

MAS-ML in order to capture the reactive agents,

proactive agents with planning and proactive agents

based on utility functions. The paper is structured as

follows. The main internal architectures for agents are

described in Section 2. Section 3 briefly presents MAS-

ML modelling language. The extension of MAS-ML is

then detailed in Section 4. In Section 5 the modelling of

the TAC-SCM (Trading Agent Competition - Supply

Chain Management) (Sadeh et al., 2003) application is

presented by using the extended MAS-ML. Related

works are described in Section 6 and, finally,

conclusions and future works are discussed in Section 7.

2 AGENT ARCHITECTURES

The agent internal architectures can be categorized

based on proactive and reactive foundations.

2.1 Simple Reflex Agents

A simple reflex (or reactive) agent (Russell and

Norvig, 2003), is considered the most simple internal

architecture. Condition-action rules are used to select

the actions based on the current perception. These

rules follow the form: “if condition then action”, and

determine the action to be executed if the perception

occurs. This architecture assumes that at any time the

agent receives information from the environment

though sensors. These perceptions consist of the

representation of state aspects that are used by the

agent for making decision. A subsystem is the one

responsible for the making decisions, i.e., responsible

for processing the perception sequence and selecting

a sequence of actions from the set of possible actions

for the agent. The agent performs the selected action

upon an environment through actuators.

2.2 Model-based Reflex Agents

The structure of this kind of agent is similar to the

simple reactive agent presented before since it deals

with the information by using condition-action rules.

In order to handle partially observable environment

and to reach a more rational performance, the agent is

able to store its current state in an internal model.

According to Weiss (1999), reflex agents with

internal states select actions by using the information

in its internal states. A function called next function is

introduced to map the perceptions and the current

internal state into a new internal state used to select the

next action. Such state describes aspects of the world

(called model) that cannot be seen in the current

moment, but it was perceived previously or has come

out by inferences (Russell and Norvig, 2003).

2.3 Goal-based Agents

Sometimes, the knowledge about the current state of

the environment is not enough to determine the next

action and additional information about desirable

situations is required. Goal-based agents are model-

based agents that set a specific goal and select the

actions that lead to that goal. This allows the agent to

choose a goal state among multiple possibilities.

Planning activity is devoted to find the sequence

of actions that are able to achieve the agent's goals

(Russell and Norvig, 2003). The sequence of actions

previously established leads the agent to reach a goal

is termed plan (Silva, and Lucena, 2004) (Silva,

Choren and Lucena, 2008a). Thus, the goal-based

agent with planning involves the next function

component and also includes the following elements:

• Formulate Goal Function, which receives the state

and returns the formulated goal.

• Formulate Problem Function, which receives the

state and the goal and returns the problem.

• Planning, that receives the problem and uses search

and/or logic approach to find a sequence of actions

to achieve a goal.

• Action that is represented with its pre-conditions

and post-conditions.

2.4 Utility-based Agents

Considering the existence of multiple goal states, it is

possible to define a measure of how desirable a

particular state is. In this case, aiming to optimize the

agent performance, the utility function is responsible

for mapping a possible state (or group of states) to a

measure of utility associated, according to the current

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

goals (Russell and Norvig, 2003). Thus, the utility

function is incorporated into the architecture.

In addition, the utility-based agent preserves the

same elements that a goal-based agent: Next function,

formulate goal function, formulate problem function,

planning and action.

3 MAS-ML

MAS-ML was originally designed to support the

modelling of proactive agents that are goal-based and

guide by pre-established plans.

MAS-ML model all structural and dynamic aspects

defined in TAO metamodel by extending the UML

metamodel. The structural diagrams defined by MAS-

ML are Role Diagram, Class Diagram and Organization

diagram (Silva, Choren and Lucena, 2004). Using the

three static diagrams is possible model all structural

aspects of entities defined in TAO. The main element

of the agent-oriented modelling is the agent itself.

Figure 1 shows the diagram element used in static

diagrams of MAS-ML to represents agents.

Figure 1: An instance of AgentClass metaclasse.

The dynamic diagrams defined in MAS-ML are

extended versions of the UML Sequence Diagram and

Activities Diagram (Silva, Lucena and Choren, 2008b).

4 MAS-ML EXTENSIONS

This section presents the extension to MAS-ML in

order to support the modelling of agents by using

diverse internal architectures: Simple reflex, Model-

based reflex, Goal-based and Utility-based. The new

version of MAS-ML is named of MAS-ML 2.0.

According to UML (2009), tagged values,

stereotypes and constraints are extension mechanisms.

Additionally, adaptation of existing metaclasses and

definition of new metaclasses can also be used.

Stereotypes and definition of new metaclasses was

used to represent Simple reflex agents, Model-based

reflex agents, Goal-based agents with planning and

Utility-based agents. Following the architecture

definitions presented in Section 2, the characteristics that

need to be defined are Perception, Next-function,

Formulate-goal-function, Formulate-problem-function,

Planning and Utility-function. The Figure 2 illustrates

the MAS-ML metamodel extensions.

Figure 2: MAS-ML metamodel extension.

The Perception collects information about the

environment and/or other agents, without modify

them. Since there is not any metaclasse in MAS-ML

that can be used to represent such concept, the

AgentPerceptionFunction metaclass was created to

represent the agent perception.

The agent perceptions can be also represented on

the environment, since it represents the elements that

the agent can perceives and as far as the sensors of

the agent will perceives (partially or fully, for

example). Once the environment influences the

perception of the therein agents, an association was

established between the metaclasses

AgentPerceptionFunction and EnvironmentClass.

The Planning task results in a sequence of actions

in order to achieve a goal (Russell and Norvig, 2003).

In addition, the following properties are observed: (i)

unlike a plan (represented by AgentPlan in the original

MAS-ML metamodel), the sequence of actions is

created at runtime; and (ii) unlike a simple action

(represented by AgentAction in the MAS-ML

metamodel), the action of planning has a goal

associated. Thus, the new metaclass

AgentPlanningStrategy was created to represent the

planning. An association relationship between

AgentPlanningStrategy and AgentPlan was defined to

represent that the action of planning can create plans.

The metaclasses AgentPerceptionFunction and

AgentPlanningStrategy extends the BehavioralFeature

metaclass. The AgentPerceptionFunction has a

Constraint that is used to restrict the information that

can be perceived through the agent sensors.

The Next-function, Formulate-goal-function,

Formulate-problem-function and Utility-function are

special agent actions that depend on the agent internal

architecture. The <<next-function>>, <<formulate-

goal-function>>, <<formulate-problem-function>> and

EXTENDING MAS-ML TO MODEL PROACTIVE AND REACTIVE SOTWARE AGENTS

<<utility-function>> stereotypes was thus created and

related to AgentAction metaclass.

Finally, the condition action rules of reactive

agents, can be represented by using the agent's action

representation (Silva, Choren and Lucena, 2008a),

which may have a pre-condition attached.

4.1 Static Representation of AgentClass

The new structural and behavioural features in the

modelling of the different types of agents influence the

AgentClass metaclass representation in static diagrams.

4.1.1 Simple Reflex Agent Structure

The representation for a Simple Reflex Agent (Figure

13) does not include any structural element since

neither goals nor beliefs are inherent to this

architecture. In the lower compartment the perceptions

and actions, driven by condition-action rules and not

by a specific plan, are represented.

4.1.2 Model-based Reflex Agent Structure

The model-based reflex agents represent an upgrade

over the simple reflex agents. Thus, the definition for

the action element is kept the same. In addition, beliefs

representing the state and the next function are

included. Figure 12 presents the graphical representation

of AgentClass for a Model-based reflex agent.

4.1.3 Goal-based Agent with Plan Structure

The goal-based agents with plan have the same structure

proposed initially by Silva and Lucena (2004) including

goals, beliefs, actions and plan. Figure 1 and Figure

11 show the graphical representation of this agent.

4.1.4 Goal-based Agent with Planning

Structure

The goal-based agents with planning incorporate

additional complexity to the agent representation.

Firstly, goals are considered in order to guide the agent

behaviour. Thus, this element is included as a

structural component denoted with the <<goal>>

stereotype. In order to manipulate consistently goals

and states, the agent behaviour is enhanced with

<<perceives>>, <<formulate-goal-function>> and

<<formulate-problem-function>> elements. The already

existent <<next-function>> element is keep up. This

function receives the current perception and the beliefs

that must be updated (state).

In addition, instead of representing pre-established

plans, the planning activity is incorporated. This

activity involves a goal and uses the available actions

to create a sequence of actions. Figure 14 illustrates the

AgentClass for a goal-based agent using planning.

4.1.5 Utility-based Agent Structure

The representation for the utility-based agent consists

in a specialization of the goal-based agent with

planning. However, the <<utility-function>> element

is added to represent the function responsible for the

optimization of the gent performance. The graphical

representation of AgentClass for a pro-active agent

based on utility is illustrated in Figure 15.

Along the planning, the agents may be linked to

reach more than one goal. In this case, the occurrence

of conflicting goals or the existence of several states

meeting the goals is possible. So the utility function is

incorporated into the agent structure, in order to

evaluate the usefulness degree of the associated goals.

4.2 AgentRoleClass Static

Representation

An AgentRoleClass in MAS-ML is represented by a

solid rectangle with a curve at the bottom. Similarly to

the class representation, it has three compartments

separated by horizontal lines. The upper compartment

contains the agent role name unique in its namespace.

The intermediate compartment contains a list of goals

and beliefs associated with the role, and below, a list of

duties, rights and protocols.

Reactive agents have not explicit goals and, more

particularly, the simple reflex agents do not have

beliefs. Thus, their role representation must be adapted

(Figure 16). In addition to the representation of the

roles of simple reflex agents, roles for model-based

reflex agents include beliefs in order to partially handle

observable environments. The agent role representation

in this case is represented in Figure 17.

The features of agent roles in other architectures

are inaltered since both define beliefs and goals. The

structural changes regarding the AgentRoleClass entity

impact the Organization and Roles diagrams.

4.3 Dynamic representation

of AgentClass

Similarly to the static diagrams, the new representation

of the AgentClass influences the representation of their

behavioural features. In follows, the dynamic

representation of the different agent types is illustrated

through sequence diagrams using MAS-ML 2.0.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

4.3.1 Reflex Agents Representation

The agent’s perception is represented in sequence

diagram of MAS-ML by an arrow with an open mind

leaving the agent to the environment, together of the

<<perceives>> stereotype, the perception name and the

elements that agent can see (Figure 3).

Figure 3: Perception of the agent in sequence diagram.

The actions sequence of the reactive agents cannot

be defined in the analysis phase, but it is possible to

represent the set of actions with the condition or

conditions associated. Figure 4 illustrates the action

taken by a reactive agent.

Figure 4: Reactive agent action in the sequence diagram.

The next function is represented in the sequence

diagram of MAS-ML 2.0 by a closed arrow with full

head, which starts at the agent and ends at the agent,

then the stereotype <<next-function>> followed by

the name of the function. Figure 5 illustrates the next

function in the sequence diagram.

Figure 5: Sequence diagram of the next function.

Therefore, if a simple reflex agent is modelled

initially have their perception and then their actions

guided by the condition-action rules. In the case of a

Model-based reflex agent, initially we have the

perception, then the next function and, finally, its

actions guided by the condition-action rules.

4.3.2 Proactive Agent Representation

The next function is executed before the formulate goal

function and is used by two types of pro-active agents

in this paper. The next function element is represented

in the sequence diagram by a arrow full head, which

begins and ends on the agent in itself, together with the

stereotype <<next-function>> and the function name

as shown in Figure 5. Then we have the representation

of the formulate problem function in the sequence

diagram. The representation is done by an arrow full

head, which begins in the agent and ends in itself,

accompanied by the relevant stereotype. The figures 6

and 7 illustrate the formulate goal function and the

formulate problem function, respectively.

Figure 6: Formulate goal function in sequence diagram.

Figure 7: Formulate problem function in sequence diagram.

In case of agents that use planning, the actions

sequence that the agent will take to achieve the goal

cannot be advanced before its execution. In this case,

planning is represented by a closed arrow head that

begins and ends in the agent in itself accompanied by the

stereotype <<planning>>. The actions that can be used

for planning to achieve (s) objective (s) are represented

as initially defined by Silva (2004). Optionally, the

criterion or algorithm used to perform the planning can

be specified by a textual note. Figure 8 illustrates the

planning in the sequence diagram of MAS-ML 2.0.

Figure 8: Planning in the sequence diagram.

The utility function element is represented in the

sequence diagram by an arrow with full head that

begins in the agent and ends in itself, together with

the stereotype <<utility-function>>. Figure 9

illustrates the representation of the utility function in

the sequence diagram of MAS-ML 2.0.

Figure 9: Utility-function in the sequence diagram.

The agent actions in MAS-ML 2.0 are modelled

using the iteration element and combined fragment,

existing in UML. This representation allows the

modelling of any combination of actions. An example is

shown in Figure 10. Since the sequence of actions for

the agents with planning is generated at runtime, the

modelling of this sequence is not required.

The agent goal-based with plan maintains the

representation proposed by Silva (2004), as well as

the plan defined during the design phase.

In the case of agent goal-based with planning,

initially runs perception, next function, formulate goal

EXTENDING MAS-ML TO MODEL PROACTIVE AND REACTIVE SOTWARE AGENTS

function, formulate problem function and then

execution of its planning, which results in the

execution of possible actions associated with the agent.

Figure 10: Implementation of the actions of the agent with

planning the sequence diagram of MAS-ML 2.0.

Finally, the agent-based utility needs the

perception, next function, formulate goal function,

formulate problem function, planning, utility function

and results in actions that are performed in that order.

4.4 Dynamic Representation

of AgentClass

The features proposed by Silva, Choren and Lucena

(2005) for the activity diagram, were maintained. Thus,

each activity is represented by a rounded rectangle.

The agent beliefs are represented by a square with the

identification of the agent used by the beliefs and goals

in the upper right corner through a textual description

denoted with the <<Goal>> stereotype.

4.4.1 Reflex Agent Representation

The activity diagram of simple and model-based reflex

agent represents the behavior from perception to

action. The behavior of a simple reflex agent is

represented on the activity of MAS-ML 2.0 as follows:

the initial activity is the perception of the agent, on the

basis of the current perception the condition action

rules are used to select one of the possible actions.

Finally, the selected action is performed.

In another hand, the behavior of a model-based

reflex agent is represented on the activity of MAS-ML

2.0 as follows: the initial activity is the perception of

the agent that can be used by the next function to

update its beliefs. After that, the condition-action rules

are responsible to select one of the possible actions.

Finally the selected action is performed.

4.4.2 Proactive Agent Representation

The activity diagram of the goal-based agent with

planning represents the agent behaviour from

perception to action. The behavior of a goal-based

agent is represented on the activity of MAS-ML 2.0

as follows: the initial activity is the perception of the

agent, after that the next function updates the beliefs

based on current perception. The formulate goal and

the formulate problem functions are executed. The

planning is performed to determine the action(s)

should be taken. Finally, the selected action(s) is

performed.

The behaviour of utility-based agent is represented

on the activity of MAS-ML 2.0 as follows: the initial

activity is the perception, then, the next function updates

beliefs based on current perception. The formulate goal

function and the formulate problem function are

executed. The planning is performed to determine what

action should be taken. The utility function helps the

choice of action, and the selected actions are performed.

5 CASE STUDY

A TAC-SCM application is used to illustrate the use of

MAS-ML 2.0 where agents with different architectures

are elicited to model different strategies in the problem.

5.1 TAC-SCM

TAC (Trading Agent Competition) (Wellman et al.,

2002) is an environment that enables the achievement

of simultaneous auctions, test techniques, algorithms

and heuristics to use in negotiation. There are two

types of games in competition: TAC-Classic (Wellman

et al., 2002) and TAC-SCM (Sadeh et al., 2003).

The TAC-SCM is concerned in planning and

managing the organization activities across a supply

chain. The TAC-SCM scenario is designed to capture

the challenges in an integrated environment for

acquisition of raw materials, production and delivery of

finished goods to customers. This environment is highly

dynamic, stochastic and strategic (Arunachalam, 2004).

The game starts when one or more agents connect

to a server game. The server simulates suppliers and

customers, providing a bank, manufacturing and

service of storage of goods to individual agents. The

game occurs along on a fixed number of simulated

days, and in the end, the agent with largest sum of

money in the bank is the winner (Collins et al., 2006).

5.2 Modelling TAC-SCM

with MAS-ML

The internal architecture of each agent in TAC-SCM is

elected according to the function in the game.

The DeliveryAgent needs to satisfy the goal of

delivery products to customers. In order to achieve this

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

goal a sequence of actions must be executed. Thus the

representation of the DeliveryAgent is modelled by

using a goal-based with plan architecture (Figure 11).

Figure 11: A DeliveryAgent proposed to TAC-SCM.

The SellerAgent offers computers to customers

and gets the payment, and the BuyerAgent decides

when to make new request for quote and realizes the

payment. Since, reactive agents reply quickly to the

perceptions (Weiss, 1999), the BuyerAgent and

SellerAgent are modelled as reactive agents aiming

the necessity of a fast reply in auction. Figure 12 and

13 shows the BuyerAgent (Model-based reflex agent)

and SellerAgent (Simple reflex agent), respectively.

Figure 12: A BuyerAgent proposed to TAC-SCM.

Figure 13: A SellerAgent proposed to TAC-SCM.

The ProductionAgent needs to satisfy current

demand across the assembling of computers and

management of the stock. To objectify achieve this goal,

it can’t use a pre-established plan because this dynamic

scenario requires a different set of actions depending on

the current demand. Thus the ProductionAgent is a goal-

based with planning agent detailed in Figure 14.

Figure 14: A ProductionAgent proposed to TAC-SCM.

Finally, the ManagerAgent is incumbed for manage

all agents and the allocation resources. This agent tries

to maximize gain and sales. Note that its goals can be

in conflict. Thus, the most appropriate architecture in

this case is the Utility-based architecture (Figure 15).

Figure 15: A ManagerAgent proposed to TAC-SCM.

The roles of reactive agents: SellerAgent and

BuyerAgent are illustrated in the figures 16 and 17,

respectively. The roles for proative agents are not

represented since they are not affected.

Figure 16: Role of SellerAgent proposed to TAC-SCM.

Figure 17: Role of BuyerAgent proposed to TAC-SCM.

Finally, Figure 18 depicts the Organization

Diagram for TAC-SCM MAS. This diagram

represents the TacOrganization and describes the

agents and agent roles in the specific environment.

Figure 18: Organization Diagram proposed to TAC-SCM.

The activity diagrams in figures 19, 20, 21, 22 and

23 describe the behavior of each agent role.

Figure 19: Role of DeliveryAgent proposed.

EXTENDING MAS-ML TO MODEL PROACTIVE AND REACTIVE SOTWARE AGENTS

Figure 20: Role of SellerAgent proposed to TAC-SCM.

Figure 21: Role of BuyerAgent proposed.

Figure 22: Role of ProductionAgent proposed.

Figure 23: Role of ManagerAgent proposed.

Figure 24 describes the sequence diagram of

DeliveryAgent. Note that the actions taken by the agent

are guided by a plan, so it is a sequence of actions.

Figure 24: Sequence Diagram of DeliveryAgent.

The sequence diagram of the SellerAgent (Figure

25) shows their execution through its perceptions and a

set of actions associated with a condition-action rule.

Figure 25: Sequence Diagram of SellerAgent.

The sequence diagram of the BuyerAgent (Figure

26) illustrates the agent perception, next function and a

set of actions associated with a condition-action rule.

Figure 26: Sequence Diagram of BuyerAgent.

Figure 27 shows the sequence diagram of the

ProductionAgent. Note that the actions taken by the

agent are result of the perception, next function,

formulate goal function, formulate problem function and

planning. Moreover, its actions are represented by a set

of possible actions.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

Figure 27: Sequence Diagram of ProductionAgent.

In Figure 28 is presented the sequence diagram of

ManagerAgent. In this case, the actions of the

ManagerAgent are guide by perception, next

function, formulate goal function, formulate problem

function, planning and utility function. Moreover, its

actions are represented by a set of possible actions.

Figure 28: Sequence Diagram of ManagerAgent.

6 RELATED WORKS

Several languages have been proposed for the

modelling of MAS. However, it does not support the

modelling of different internal architectures of agents

available in Russell and Norvig (2003) and Weiss

(1999). Besides, they have several drawbacks that have

justified the choosing of MAS-ML to be extended in

order to model different agent architectures.

The work of Odell, Parunak and Bauer (2000)

presents the AUML language. This modelling

language aims to provide a semi-formal and intuitive

semantics through a friendly graphical notation.

AUML does not provide elements to represent

perceptions and the next-function.

Wagner (2003) proposes the AORML modelling

language, which is based on the AOR metamodel. This

language does not give support to modelling of the

elements of the internal agent architectures. Therefore

it is not possible to differentiate agents with reactive

and proactive architectures in AORML.

Moreover, the two languages mentioned above do

not define the environment as an abstraction, so it is

not possible to model the agent migration from an

environment to another. This capability is inherent to

mobile agents modelling (Silva. and Mendes, 2003).

Choren and Lucena (2004) present the ANote

modelling language, involving a set of models, called

views. In ANote, it is not possible to differentiate

agents with reactive architectures from the proactive

ones. In addition, ANote does not support conventional

objects, used to model non-autonomous entities. The

language defines several concepts related to agents, but

the concept of agent role is not specified. This concept

is extremely important when modelling societies where

agents can play different roles at the same time.

AML (Cervenka et al., 2004) is a modelling

language based on a metamodel that enables the

modelling of organizational units, social relations,

roles and role properties. AML gives adequate support

for the modelling of reactive agents, goal-based agents

with planning and utility-agents. It is worth mentioning

though the semantic aspects of communication are

modelled as specializations of existing elements in

UML, such as methods invocation, what is not

adequate for modelling agent communication.

7 CONCLUSIONS

This paper presents an extension to MAS-ML language

in order to allow the modelling of diverse internal

agent architectures published in the agent literature,

such as: Simple reflex agents, Model-based reflex

agents, Goal-based agents and Utility-based agents.

EXTENDING MAS-ML TO MODEL PROACTIVE AND REACTIVE SOTWARE AGENTS

MAS-ML was originally designed to support the

modelling of pro-active goal-based agents with plan.

Thus, some issues were detected while trying to use the

language to model reactive agents and other pro-active

architectures. In this sense, the MAS-ML evolution

proposed in this work involves the definition of two

new metaclasses AgentPerceptionFunction and

AgentPlanningStrategy in order to aggregate the

representation of different agent behaviour. Also, new

stereotypes to describe the behaviour of agent from

specific architectures were defined and associated to

AgentAction metaclass. The static structure of

AgentClass and AgentRoleClass entities were also

modified. Then, the class, organization, role, sequence

and activity diagrams were changed in consistency.

The modelling tool to support the proposed

approach is also under development. Other case studies

are being conducted to provide further validation to

this work. Moreover, the possibility of MAS-ML

extension for other internal architectures, such as the

BDI architecture is an interesting possibility.

REFERENCES

Arunachalam, R, 2004. The 2003 supply chain management

trading agent competition. In: Third International Joint

Conference on Autonomous Agents & Multi Agent

Systems. July 2004. [S.l.: s.n.]. p. 113–120.

Cervenka, R., Trencansky, I., Calisti, M., and Greenwood, D

2004. AML: Agent Modeling Language Toward

Industry-Grade Agent-Based Modeling. In: Agent-

Oriented Software Engineering, pp. 31–46. Springer-

Verlag, Berlin.

Choren, R. and Lucena, C., 2004. "Agent-Oriented

Modeling Using ANote", 3rd International Workshop

on Software Engineering for Large-Scale Multi-Agent

Systems. The Institution of Electrical Engineers, IEE,

Stevenage, UK, 2004, pp. 74-80.

Collins, J.; Arunachalam, R.; Sadeh, N.; Eriksson, J.; Finne,

N.; Janson, S., 2006. The Supply Chain Management

Game for the 2007 Trading Agent Competition.

Available in http://www.sics.se/tac/tac07scmspec.pdf.

Dorigo, M. and Stützle, T., 2004. Ant Colony Optimization.

The MIT Press, Cambridge, Massachusetts.

Jennings, Nicholas R., 1996. Coordination Techniques for

Distributed Artificial Intelligence. In: Foundations of

Distributed Artificial Intelligence, pp. 187-210, Wiley.

Odell, J., Parunak, H. V. D., Bauer, B., 2000. Extending

UML for Agents. In Agent-Oriented Information

Systems Workshop at the 17th National conference on

Artificial Intelligence (AIII’00) (3-17).

Russell, S. and Norvig, P., 2003. Artificial Intelligence: A

Modern Approach, 2nd Ed., Upper Saddle River, NJ:

Prentice Hall, ISBN 0-13-790395-2,

Sadeh, N.; Arunachalam, R.; Erikson, J.; Finne, N.; Janson,

S., 2003. A supply-chain trading competition. AI

Magazine, v. 24, n. 1, p. 92–94

Silva, P. S. e Mendes, M. J. (2003). Uma Abordagem para

Incorporar Mecanismos de Inteligência Artificial a

Agentes Móveis. XXI Simpósio Brasileiro de Redes de

Computadores. Natal, Rio Grande do Norte.pp 837-

852.

Silva, V.; Lucena, C. 2004. From a Conceptual Framework

for Agents and Objects to a Multi-Agent System

Modeling Language, In Journal of Autonomous Agents

and Multi-Agent Systems, volume 9, issue 1-2, Kluwer

Academic Publishers, pp. 145-189, 2004.

Silva, V.; Choren, R.; Lucena, C. 2004. A UML Based

Approach for Modeling and Implementing Multi-Agent

Systems. In: Proceeding of the third International

Conference on Autonomous Agents and Multi-Agents

Systems. New York, USA, IEEE Computer Society,

volume 2, pp. 914-921.

Silva, V. T. da, Choren, R., Lucena, C. J. P. de (2005).

Using UML 2.0 Activity Diagram to Model Agent

Plans and Actions. In: 4th International Conference on

Autonomous Agents and Multi-Agent Systems

(AAMAS), Netherlands, pp. 594-600.

Silva, V.; Choren R.; Lucena, C. 2008 a. MAS-ML: A

Multi-Agent System Modelling Language, In

International Journal of Agent-Oriented Software

Engineering, Interscience Publishers, vol.2, no.4.

Silva, V.; Choren, R.; Lucena, C. 2008 b. Modeling MAS

Properties with MAS-ML Dynamic Diagrams. In 8th

International Bi-Conference Workshop, LNCS 4898,

Springer-Verlag, pp. 1-18.

Wagner, G., 2003. The Agent-Object-Relationship Meta-

Model: Towards a Unified View of State and Behavior.

Information Systems, v. 28, n.5, pp. 475–504.

Weiss, G., 1999. Multiagent Systems: A Modern Approach

to Distributed Artificial Intelligence. MIT Press,

Massachusetts.

Wellman, M. P.; Stone, P.; Greenwald, A.; Wurman, P. R.,

2002. The 2001 Trading Agent Competition. IEEE

Internet Computing, v. 13, p. 935–941.

UML: Unified Modeling Language Specification, version

2.2, OMG, 2009 available in: http://

www.omg.org/technology/documents/modeling_spec_c

atalog.htm#UML.

Zambonelli, F.; Jennings, N.; Wooldridge, M., 2001.

Organizational abstractions for the analysis and design

of multi-agent systems. In: Agent-Oriented Software

Engineering, LNCS 1957, Berlin: Springer, p. 127-141.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems