Multi-Agent Systems: Theory and Application in

Organization Modelling

Joaquim Filipe

Escola Superior de Tecnologia do Instituto Politécnico de Setúbal

Rua Vale de Chaves, Estefanilha, 2910 Setúbal, Portugal

joaquim.filipe@estsetubal.ips.pt

1 Introduction

Organisations are multi-agent systems, eventually including both human and artificial

agents. Organisations are also seen as multilayered Information Systems (IS)

themselves, including an informal subsystem, a formal subsystem and a technical

system as shown in figure 1.

Fig. 1. Three main layers of the real information system [10].

We aim at improving the technical subsystem within the constraints defined by the

other two. Organisational information systems are inherently distributed, nowadays,

thus communication and coordination are major problems in this kind of information

systems. Perhaps motivated by the difficult problems there is currently a strong

interest on this area, which is an active research field for several disciplines, including

Distributed Artificial Intelligence (DAI), Organisational Semiotics and the Language-

Action Perspective, among others. Our approach integrates elements from these three

perspectives.

The Epistemic-Deontic-Axiologic (EDA) designation refers to the three main

components of the agent structure described in this paper. Here we propose an agent

model which, contrary to most DAI proposals, not only accounts for intentionality but

is also prepared for social interaction in a multi-agent setting. Existing agent models

emphasise an intentional notion of agency – the supposition that agents should be

INFORMAL IS: a sub-culture where meanings are established, intentions are understood,

beliefs are formed and commitments with responsibilities are made, altered and discharged

FORMAL IS: bureaucracy where form and rule replace meaning and intention

TECHNICAL IS: Mechanisms to automate part of the formal

system

Filipe J.

Multi-Agent Systems: Theory and Application in Organization Modelling.

DOI: 10.5220/0004466200030014

In Proceedings of the 4th International Workshop on Enterprise Systems and Technology (I-WEST 2010), pages 3-14

ISBN: 978-989-8425-44-7

understood primarily in terms of mental concepts such as beliefs, desires and

intentions. The BDI model [9] is a paradigm of this kind of agent.

We claim that the cognitive notions that constitute the basis of intentional models

show only one face of the coin, the other one being social agency. It is required that

an adequate agent model, able to function in a multi-agent setting, should emphasise

social agency and co-ordination, within a semiotics framework. Our approach focuses

on organisational agents who participate in organisational processes involving the

creation and exchange of signs, i.e. knowledge sharing, in a way that is inherently

public, thus depending on the agent’s social context, i.e. its information field [10], to

support co-ordinated behaviour. An information field is composed by all the agents

and objects involved in a certain activity and sharing a common ontology.

A realistic social model must be normative: both human agents and correctly

designed artificial agents ought to comply with the different kinds of norms that

define the information field where they operate, although exceptions may occur.

Private representations of this shared, normative, knowledge are translations of the

public knowledge into specific agent mental structures. When designing an artificial

agent, the designer must adopt a private knowledge representation paradigm to set up

the agent’s internal knowledge in such a way that it fits the normative shared

ontology.

We postulate that norms are the basic building blocks upon which it is possible to

build co-ordination among organised entities and co-ordinated actions are the crux of

organised behaviour. We claim that although organised activity is possible either with

or without communication, it is not possible without a shared set of norms. Therefore,

these socially shared norms define an information field that is a necessary condition

for heterogeneous multi-agent co-ordination including both artificial agents and

humans.

2 The Normative Structure of the EDA Model

Norms are typically social phenomena. This is not only because they stem from the

existence of some community but also because norms are multi-agent objects [5]:

• They concern more than one individual (the information field involves a

community)

• They express someone’s desires and assign tasks to someone else

• Norms may be regarded from different points of view, deriving from their

social circulation: in each case a norm plays a different cognitive role, be it a

simple belief, a goal, a value, or something else.

Social psychology provides a well-known classification of norms, partitioning them

into perceptual, evaluative, cognitive and behavioural norms. These four types of

norms are associated with four distinct attitudes, respectively [10]:

• Ontological – to acknowledge the existence of something;

• Axiologic – to be disposed in favour or against something in value terms;

• Epistemic – to adopt a degree of belief or disbelief;

• Deontic – to be disposed to act in some way.

Fig. 2. The EDA Agent Knowledge-Base Structure.

An EDA agent is a knowledge-based system whose knowledge base structure is

based on the following three components: the Epistemic, the Deontic and the

Axiologic.

The epistemic model component is where the knowledge of the agent is stored, in

the form of statements that are accepted by that agent. Two types of knowledge are

stored here: declarative knowledge – statements relative to the agent beliefs – and

procedural knowledge – statements concerning the know-how of the agent, e.g. their

plans and procedural abilities.

The importance of norms to action has determined the name we have chosen for

the model component where the agent goals are represented. An agent goal may be

simply understood as the desire to perform an action (which would motivate the

designation of conative) but it can also be understood, especially in a social context,

as the result of the internalisation of a duty or social obligation (which would

motivate the designation of deontic). We have adopted the latter designation not only

because we want to emphasise the importance of social obligations but also because

personal desires can be seen as a form of ‘generalised’ obligation established by an

agent for himself. This provides a unification of social and individual drives for

action, which simplifies many aspects of the model.

The axiologic model component contains the value system of the agent, namely a

partial order that defines the agent preferences with respect to norms. The importance

of the agent value system is apparent in situations of conflict, when it is necessary to

violate a norm. This preference ordering is dynamic, in the sense that it may change

whenever the other internal components of the agent model change, reflecting

different beliefs or different goals.

3 Intentions and Social Norms in the EDA Model

The multi-agent system metaphor that we have adopted for modelling organisations

implies that organisations are seen as goal-governed collective agents, which are

composed of individual agents.

In our model individual agents are autonomous, heterogeneous, rational, social

agents. Therefore, they are compelled to make decisions and act in a way that,

although not entirely deterministic, is constrained by their rationality. Actually their

behaviour would be predictable if we knew all the details of their EDA model

EDA Agent Knowledge-Base

Epistemic

Component

Deontic

Component

Axiologic

Component

components, the environment stimuli, their perception function and also the reasoning

machine they use, because the ultimate goal of a rational agent is to maximize its

utility.

Typically, artificial intelligence (AI) agent models consider intelligent agent

decision processes as internal processes that occur in the mind and involve

exclusively logical reasoning, external inputs being essentially data that are perceived

directly by the agent. This perspective does not acknowledge any social environment

whatsoever. In this paper we start from a totally different perspective, by emphasising

the importance of social influences and a shared ontology on the agent decision

processes, which then determines agents’ activity. We shall henceforth refer to agent

as ‘it’ although the EDA model also applies to human agents. In any case, we are

particularly interested in the situations where information systems are formally

described, thus making it possible for artificial agents to assist or replace human

agents.

An important role of norms in agent decision processes is the so-called cognitive

economy: by following norms the agent does not have to follow long reasoning chains

to calculate utilities – it just needs to follow the norms.

However, instead of adopting a whole-hearted social sciences perspective, which

is often concerned merely with a macro perspective and a statistical view of social

activity, we have adopted an intermediate perspective, where social notions are

introduced to complement the individualistic traditional AI decision models: a

psycho-social perspective, whereby an agent is endowed with the capability of

overriding social norms by intentionally deciding so.

Our model enables the relationship between socially shared beliefs with agent

individual, private, beliefs; it also enables the analysis of the mutual relationships

between moral values at the social level with ethical values at the individual level.

However, we have found particularly interesting analogies in the deontic component,

specifically in the nature of the entities and processes that are involved in agent goal-

directed behaviour, by inspecting and comparing both the social processes and

individual processes enacted in the deontic component of the EDA model.

This was motivated by the close relationship between deontic concepts and agency

concepts, and represents a direction of research that studies agency in terms of

normative social concepts: obligations, responsibilities, commitments and duties.

These concepts, together with the concepts of power/influence, contribute to facilitate

the creation of organisational models, and are compatible with a vision of

organisations as normative information systems as well as with the notion of

information field that underlies the organisational semiotics approach, on which the

work presented in this paper is inspired.

As will be explained in more detail in the next section, an essential aspect of the

EDA model is that the Deontic component is based on the notion of generalised goal

as a kind of obligation, that encompasses both social goals (social obligations) and

individual goals (self obligations). Following a traditional designation in DAI, we

designate those individual generalised goals that are inserted in the agenda as

achievement goals, as in [4]. Figure 3 describes the parallelism between mental and

social constructs that lead to setting a goal in the agenda, and which justifies the

adoption of the aforementioned generalised obligation. Here, p represents a

proposition (world state).

()Bp

represents p as one of agent

’s beliefs.

()Op

Fig. 3. Social and Individual goals parallelism in the EDA model.

represents the obligation that

must see to it that p is true for

()Op

represents

the interest that

has on seeing to it that p is true for itself – a kind of self-imposed

obligation. In this diagram

(,)

∈Ε

means, intuitively, that proposition p is

one of the goals on

’s agenda.

Interest is one of the key notions that are represented in the EDA model, based on

the combination of the deontic operator ‘ought-to-be’ [14] and the agentive ‘see-to-it-

that’ stit operator [1]. Interests and Desires are manifestations of Individual Goals.

The differences between them are the following:

• Interests are individual goals of which the agent is not necessarily aware,

typically at a high abstraction level, which would contribute to improve its

overall utility. Interests may be originated externally, by other agents’

suggestions, or internally, by inference: deductively (means-end analysis),

inductively or abductively. One of the most difficult tasks for an agent is to

become aware of its interest areas because there are too many potentially

advantageous world states, making the full utility evaluation of each potential

interest impossible, given the limited reasoning capacity of any agent.

• Desires are interests that the agent is aware of. However, they may not be

achievable and may even conflict with other agent goals; the logical

translation indicated in the figure,

() ( ())Op BOp

αα

ααα

∧

, means that desires

are goals that agent

ought to pursue for itself and that it is aware of.

However, the agent has not yet decided to commit to it, in a global

perspective, i.e. considering all other possibilities. In other words, desires

Individual Goal

Social

Obligation

Interest

()Op

Duty

()Op

Origin:

- Other agents

- Inference

Origin:

- Social roles

- Inference

Desire

() ())(Op BOp

αα

∧

Demand

() ())(Op BOp

ββ

αα

∧

Intention

(,)pWD

∈Ε

Agenda

Achievement

Goals

External Internal (in EDA)

Awarene

become intentions only if they are part of the preferred extension of the

normative agent EDA model [7].

It is important to point out the strong connection between these deontic concepts

and the axiologic component. All notions indicated in the figure should be interpreted

from the agent perspective, i.e. values assigned to interests are determined by the

agent. Eventually, external agents may consider some goal (interest) as having a

positive value for the agent and yet the agent himself may decide otherwise. That is

why interests are considered here to be the set of all goals to which the agent would

assign a positive utility, but which it may not be aware of. In that case the

responsibility for the interest remains on the external agent.

Not all interests become desires but all desires are agent interests. This may seem

contradictory with a situation commonly seen in human societies of agents acting in

others’ best interests, sometimes even against their desires: that’s what parents do for

their children. However, this does not mean that the agent desires are not seen as

positive by the agent; it only shows that the agent may have a deficient axiologic

system (by its information field standards) and in that case the social group may give

other agents the right to override that agent. In the case of artificial agents such a

discrepancy would typically cause the agent to be banned from the information field

(no access to social resources) and eventually repaired or discontinued by human

supervisors, due to social pressure (e.g. software viruses).

In parallel with Interests and Desires, there are also social driving forces

converging to influence individual achievement goals, but through a different path,

based on the general notion of social obligation. Social obligations are the goals that

the social group where the agent is situated require the agent to attain. These can also

have different flavours in parallel to what we have described for individual goals.

• Duties are social goals that are attached to the particular roles that the agent is

assigned to, whether the agent is aware that they exist or not. The statement

()Op

means that agent

ought to do p on behalf of another agent

Agent

may be another individual agent or a collective agent, such as the

society to which

belongs. Besides the obligations that are explicitly

indicated in social roles, there are additional implicit obligations. These are

inferred from conditional social norms and typically depend on

circumstances. Additionally, all specific commitments that the agent may

agree to enter also become duties; however, in this case, the agent is

necessarily aware of them.

• Demands are duties that the agent is aware of

. This notion is formalised by

the following logical statement:

() ( ())Op BOp

ββ

ααα

∧

. Social demands

motivate the agent to act but they may not be achievable and may even

conflict with other agent duties; being autonomous, the agent may also decide

that, according to circumstances, it is better not to fulfil a social demand and

rather accept the corresponding sanction. Demands become intentions only if

they are part of the preferred extension of the normative agent EDA model –

see [7] section 5.7 for details.

According to the Concise Oxford Dictionary, demand is “an insistent and peremptory request,

made as of right”. We believe this is the English word with the closest semantics to what we need.

• Intentions: Whatever their origin (individual or social) intentions constitute a

non-conflicting set of goals that are believed to offer the highest possible

value for the concerned agent. Intentions are designated by some authors [11]

as psychological commitments (to act). However, intentions may eventually

(despite the agent sincerity) not actually be placed in the agenda, for several

reasons:

o They may be too abstract to become directly executed, thus requiring

further means-end analysis and planning.

o They may need to wait for their appropriate time of execution.

o They may be overridden by higher priority intentions.

o Required resources may not be ready.

The semantics of the prescriptive notions described above may be partially

captured using set relationships as depicted in figure 4, below.

Fig. 4. Set-theoretic relationships among deontic prescriptive concepts.

When an agent decides to act in order to fulfil an intention, an agenda item is

created – we adopt the designation of achievement goal. Achievement goals are

defined as in [4] as goals that are shared by individuals participating in a team that has

a joint persistent goal. Following the terminology of [4] agent

has a weak

achievement goal, relative to its motivation (which in our case corresponds to the

origin and perceived utility of that goal), to bring about the joint persistent goal

either of the following is true:

•

does not yet believe that

is true and has

being eventually true as a

goal (i.e.

has a normal achievement goal to bring about

)

•

believes that

is true, will never be true or is irrelevant (utility below the

motivation threshold), but has a goal that the status of

be mutually

believed by all team members.

However we do not adopt the notion of joint persistent goal for social co-

ordination, as proposed by Cohen and Levesque [4] because their approach has a

number of shortcomings, not only theoretical but also related to the practical

feasibility of their model, which are well documented in [12].

Interests

Desires

Duties

Demands

Intentions

Agenda

Achiev.

Goals

Fig. 5. The EDA model component relationships.

4 The EDA Model Internal Architecture

Using the social psychology taxonomy of norms, and based on the assumption that

organisational agents’ behaviour is determined by the evaluation of deontic norms,

given the agent epistemic state, with axiological norms for solving eventual interest

conflicts, we propose an intentional agent model, which is decomposed into three

main components: the epistemic, the deontic and the axiologic. Additionally there are

two external interfaces: an input (perceptual) interface, through which the agent

receives and pragmatically interprets messages from the environment and an output

(acting) external interface through which the agent acts upon the environment, namely

sending messages to other agents

A socially shared ontology is partially incorporated in an agent cognitive model

whenever it is needed, i.e. when the agent needs to perform a particular role. In this

case, beliefs are incorporated in the Epistemic component, obligations and

responsibilities are incorporated in the Deontic component and values (using a partial

order relation of importance) are incorporated in the Axiologic component – all

indexed to the particular role that the agent is to play.

Figure 5 depicts the EDA model and its component relationships.

Ψ is a pragmatic function that filters perceptions, according to the agent ontology,

using perceptual and axiologic norms, and updates one or more model

components.

∑ is an axiologic function that is used mainly in two circumstances: to help decide

which signs to perceive, and to help decide which goals to put in the agenda and

execute.

Κ is a knowledge-based component, where the agent stores its beliefs both

explicitly and implicitly, in the form of potential deductions based on logical

reasoning.

Δ is a set of plans, either explicit or implicit, the agent is interested in and may

choose to execute.

The detailed description of each component, including its internal structure, is

provided in [7]. In this paper we focus on the system behaviour. The next sections

In this paper we restrict our attention to the semiotic, symbolic, types of agent activity, ignoring

substantive physical activities.

action

perception

Axiological Component

∑

(values)

Epistemic Component

(knowledge)

Deontic Component

(behavior)

show how in EDA we specify ideal patterns of behaviour and also how we represent

and deal with non-ideal behaviours.

5 Organisational Modelling with Multi-Agent Systems using the

EDA Model

The EDA model may apply to both human and artificial agents, and is concerned with

the social nature of organisational agents:

Firstly, because it accounts for a particular mental structure (Epistemic-

Deontic-Axiologic) that is better, for our purposes, than other agent mental

structures proposed in the literature to model agent interaction. Specifically,

we intend to use it for modelling information fields where social norms

influence individual agents and are used by an agent to guide inter-subjective

communication and achieve multi-agent co-ordination.

Secondly, because the model is based on normative notions that are not only

intended to guide the agent behaviour in a socially accepted way, but also to

identify what sanctions to expect from norm violations, in order to let the

agent take decisions about its goals and actions, especially when co-

ordination is involved. The EDA model is based on the claim that multi-

agent notions such as social commitment, joint intentions, teamwork,

negotiation and social roles, would be merely metaphorical if their normative

character were not accounted for.

Given its social-enabled nature, the EDA agent notion may be used to model and

implement social activities, involving multi-agent co-ordination. However, although

the agent paradigm described in this paper is suited to model team work and joint

problem solving, the major novelty with respect to other current agent models is the

normative flavour. EDA agents are able to co-ordinate on the basis of shared norms

and social commitments. Shared norms are used both for planning and for reasoning

about what is expected from other agents but, internally, EDA agents keep an

intentional representation of their own goals, beliefs and values.

Co-ordination is based on commitments to execute requested services.

Commitments are represented not as joint intentions based on mutual beliefs, as is the

case of the Cohen-Levesque model, upon which the BDI paradigm is based, but as

first-class social concepts, at an inter-subjective level. The organisational memory

embedded in the representation of socially accepted best-practices or patterns of

behaviour and the handling of sub-ideal situations is, in our opinion, one of the main

contributions that a multi-agent system can bring about.

6 Representing Ideal and non-Ideal Patterns of Behaviour

The EDA model is a norm-based model. Norms are ultimately an external source of

action control. This assumption is reflected specially in the Deontic component of the

EDA model.

Standard Deontic Logic (SDL) represents and reasons about ideal situations only.

However, although agent behaviour is guided by deontic guidelines, in reality an

agent who always behaves in an ideal way is seldom seen. The need to overcome the

limited expressiveness of SDL, and to provide a way to represent and manipulate sub-

ideal states has been acknowledged and important work has been done in that

direction, e.g. by Dignum et al. [6] and also by Carmo and Jones [2].

Contrary to SDL, the Deontic component of the EDA model is designed to handle

sub-ideal situations. Even in non-ideal worlds, where conflicting interests and

obligations co-exist, we wish to be able to reason about agent interests and desirable

world states in such a way that the agent still is able to function coherently.

Behaviours may be represented as partial plans at different abstract levels. A goal

is a very high abstract plan, whereas a sequence of elementary actions defines a plan

at the instance level. The Deontic component is similar, in this sense, to what Werner

[13] called the agent intentional state.

However, in our model, agent decisions depend both on the available plans and a

preference relationship defined in the axiologic component. This value assignment,

which is essential for determining agent intentions, i.e. its preferred actions, can

change dynamically, either due to external events (perception) or to internal events

(inference), thus dynamically modifying the agent’s intentions.

Although our representation of ideal behaviours is based in deontic logic, we

acknowledge the existence of problems with deontic logic, partially caused by the fact

that the modal ‘ought’ operator actually collapses two operators with different

meanings, namely ‘ought-to-do’ and ‘ought-to-be’. Our solution, inspired in [8] and

[1], is to use a combination of action logic and deontic logic for representing agentive

‘ought-to-do’ statements, leaving the standard deontic operator for propositional,

declarative, statements. Agentive statements are represented as

[

]

tit Q

where

stands for an agent and Q stands for any kind of sentence (declarative or agentive).

An ‘ought-to-do’ is represented using the conventional ‘ought-to-be’ modal operator

combined with an agentive statement, yielding statements of the form

[]

:OstitQ

or, for short:

()OQ

The representation of behaviours using this kind of agentive statement has several

attractive properties, including the fact that it is a declarative representation (with all

the flexibility that it provides) and that there exists the possibility of nesting plans as

nested stits.

A plan is typically given to the agent at a very abstract level, by specifying the

goal that it ought to achieve. The agent should then be able to decompose it into

simple, executable, actions. This decomposition can be achieved by a means-ends

process. By representing plans declaratively, as behavioural norms, the process

becomes similar to backward chaining reasoning from abstract goals to more specific

ones, until executable tasks are identified, the same way as in goal oriented reasoning

in knowledge-based systems. This similarity enables the adoption of methods and

tools from that area, being especially useful the inference engine concept.

However, the Deontic component does not control the Agenda directly, i.e. it is

not responsible for setting the agent goals directly, because the prospective agent

goals – similar to desires, in the BDI model – must be analysed by the Axiologic

component first, which computes their value accordingly to an internal preference

relation, taking into account possible obligation violations.

7 Conclusions

The EDA model is a norm-based, theoretically sound, agent model that takes into

account not only the intentional aspects of agency but also the social norms that

prescribe and proscribe certain agent patterns of behaviour. The main components of

this model (Epistemic, Deontic and Axiological) have a direct relationship with the

types of norms that are proposed in the social psychology theory supporting the

model. In this paper, however, we focused our attention essentially in the Deontic

component, where the normative social aspects are more important, namely where

ideal and sub-ideal behaviours are represented.

We consider agents to be goal-governed systems. All agent goals can be

represented as obligations, encompassing both agent self-imposed obligations and

social obligations – derived from moral obligations or commitments established in the

course of their social activity.

Organizations can be seen as multi-agent systems with the EDA internal

architecture based on deontic agency notions making it easier to understand some

social aspects that in other agent models can only be modelled indirectly through

joint-goals.

References

1. Belnap, N., 1991. Backwards and Forwards in the Modal Logic of Agency. Philosophy and

Phenomenological Research, vol. 51.

2. Carmo, J. and A. Jones, 1998. Deontic Logic and Contrary to Duties. Personal copy of a

pre-print manuscript.

3. Castelfranchi, C., 1993. Commitments: from Individual Intentions to Groups and

Organisations. In Working Notes of AAAI’93 Workshop on AI and Theory of Groups and

Organisations: Conceptual and Empirical Research, pp.35-41.

4. Cohen, P. and H. Levesque, 1990. Intention is Choice with Commitment. Artificial

Intelligence, 42:213-261.

5. Conte, R., and C. Castelfranchi, 1995. Cognitive and Social Action, UCL Press, London.

6. Dignum, F., J-J Ch. Meyer and R.Wieringa, 1994. A Dynamic Logic for Reasoning About

Sub-Ideal States. In J.Breuker (Ed.), ECAI Workshop on Artificial Normative Reasoning,

pp.79-92, Amsterdam, The Netherlands.

7. Filipe, J., 2000. Normative Organisational Modelling using Intelligent Multi-Agent

Systems. Ph.D. thesis, University of Staffordshire, UK.

8. Hilpinen, R., 1971. Deontic Logic: Introductory and Systematic Readings. Reidel

Publishing Co., Dordrecht, The Netherlands.

9. Rao, A. and M. Georgeff, 1991. Modeling Rational Agents within a BDI architecture.

Proceedings of the 2nd International Conference on Principles of Knowledge

Representation and Reasoning (Nebel, Rich, and Swartout, Eds.) pp.439-449. Morgan

Kaufman, San Mateo, USA.

10. Stamper, R., 1996. Signs, Information, Norms and Systems. In Holmqvist et al. (Eds.),

Signs of Work, Semiosis and Information Processing in Organizations, Walter de Gruyter,

Berlin, New York.

11. Singh, M., 1990. Towards a Theory of Situated Know-How. 9th European Conference on

Artificial Intelligence.

12. Singh, M., 1996. Multiagent Systems as Spheres of Commitment. Proceedings of the

International Conference on Multiagent Systems (ICMAS) - Workshop on Norms,

Obligations and Conventions. Kyoto, Japan.

13. Werner, E., 1989. Cooperating Agents: A Unified Theory of Communication and Social

Structure. In Gasser and Huhns (Eds.), Distributed Artificial Intelligence, pp.3-36, Morgan

Kaufman, San Mateo, USA.

14. von Wright, G., 1951. Deontic Logic. Mind, 60, pp.1-15.