A Distributed Architecture for Norm
Management in Multi-Agent Systems
A. Garca-Camino1, J. A. Rodrguez-Aguilar1, and W. Vasconcelos2
1IIIA, Articial Intelligence Research Institute 2Dept. of Computing Science
CSIC, Spanish Research Council University of Aberdeen
Campus UAB, 08193 Bellaterra, Spain Aberdeen AB24 3UE, UK
fandres,jarg@iiia.csic.es wvasconcelos@acm.org
Abstract. Norms, that is, obligations, prohibitions and permissions,
are useful abstractions to facilitate coordination in open, heterogeneous
multi-agent systems. We observe a lack of distributed architectures and
non-centralised computational models for norms. We propose a model,
viz., normative structures, to regulate the behaviour of autonomous agents
taking part in simultaneous and possibly related activities within a multi-
agent system. This artifact allows the propagation of normative positions
(that is, the obligations, prohibitions and permissions associated to indi-
vidual agents) as a consequence of agents' actions. Within a normative
structure, con
icts may arise { one same action can be simultaneousely
forbidden and obliged/permitted. This is due to the concurrent and
dynamic nature of agents' interactions in a multi-agent system. How-
ever, ensuring con
ict freedom of normative structures at design time
is computationally intractable, and thus real-time con
ict resolution is
required: our architecture support the distributed management of nor-
mative positions, including con
ict detection and resolution.
1 Introduction
An essential characteristic of open, heterogeneous multi-agent systems (MASs)
is that agents' interactions are regulated to comply with the conventions of the
system. Norms, that is, obligations, prohibitions and permissions, can be used
to represent such conventions and hence as a means to regulate the observable
behaviour of agents [3,18]. There are many contributions on the subject of norms
from sociologists, philosophers and logicians (e.g., [10,18]). Recently, proposals
for computational realisations of normative models have been presented. Some
of them operate in a centralised manner (e.g. [5,9,13]) which creates bottlenecks
and single points-of-failure. Others (e.g. [3,12]), although distributed, aim at
the regulation of communication between agents without taking into account
that some of the normative positions (i.e., their permissions, prohibitions and
obligations) generated as a result of agent interaction may also aect other agents
not involved in the communication.
The class of MASs we envisage consists of multiple, simultaneous and pos-
sibly related agent interactions, or activities. Each agent may simultaneously
participate in several activities, and may change from one activity to another.

An agent's actions within one activity may have consequences { These are cap-
tured as normative positions that dene, in
uence or constrain the agent's future
behaviour. For instance, a buyer agent who ran out of credit may be forbidden
from making further oers, or a seller agent is obliged to deliver the goods after
closing a deal. Within a MAS normative con
icts may arise due to the dynamic
nature of the MAS and simultaneous agents' actions. A normative con
ict arises,
for instance, when an action is simultaneously prohibited and obliged. Such con-
icts ought to be identied and resolved. This analysis of con
icts can be carried
out in each activity. However, ensuring con
ict-freedom on a network of agent
conversations (or activities) at design time is computationally intractable as
shown in [7].
We propose means to handle con
icting normative positions in open and
regulated MASs in a distributed manner. In realistic settings run-time con
ict
detection and resolution is required. Hence, we require a tractable algorithm for
con
ict resolution along the lines of the one presented in [7]. The only modi-
cation required for that algorithm is that it should return a list of updates (or
normative commands), that is, the norms to be added and removed, instead of
the resulting set of norms obtained from the updates.
We need an architecture to incorporate the previously mentioned algorithm.
Among other features, we require our architecture to be distributed, regulated,
open, and heterogeneous. These features are included in other architectures such
as AMELI [3]. However, the propagation of normative positions to several agents
or to an agent not directly involved in the interaction and the resolution of
normative con
icts has not yet been addressed.
We thus propose an extension of the architecture presented in [3] fullling
these features. We extend AMELI by including a new type of agent, viz., the nor-
mative managers, also adding interaction protocols with this new type of agent,
allowing for a novel conceptual dierentiation of administrative (or \internal")
agents. Thus, the main contribution of the paper is a distributed architecture to
regulate the behaviour of autonomous agents and manage normative aspects of a
MAS, including the propagation of normative positions to dierent conversations
and the resolution of normative con
icts.
This paper is organised as follows. In Section 2 we present a scenario to
illustrate and motivate our approach. Normative structures are introduced in
Section 3. Section 4 presents our distributed architecture and, in Section 5, we
comment on related work. Finally, we draw conclusions and report on future
work in Section 6.
2 Scenario
We make use of a contract scenario in which companies come together at an
online marketplace to negotiate and sign contracts in order to get certain tasks
done. The overall transaction procedure may be organised as ve distributed
activities, represented as nodes in the diagram in Figure 1. The activities involve
dierent participants whose behaviour is coordinated through protocols.

After registering at the marketplace, clients and suppliers get together in
an activity where they negotiate the terms of their contract, i.e. actions to
be performed, prices, deadlines and other details. The client will then partici-
pate in a payment activ-
ity, verifying his credit-
worthiness and instruct-
ing his bank to transfer
the correct amount of money.
The supplier in the mean-
time will delegate to spe-
cialised employees the ac-
tions to be performed in
Exit
Registration
Payment
Work
Negotiation
Coordination Level
Fig. 1: Activity Structure of the Scenario
the work activity. Finally, agents can leave the marketplace conforming to a
predetermined exit protocol. The marketplace accountant participates in most
of the activities as a trusted provider of auditing tools.
3 Normative Structure
We address a class of MASs in which interactions are carried out via illocution-
ary speech acts [14] exchanged among participating agents, along the lines of
agent communication languages such as FIPA-ACL [6]. In these MASs, agents
interact according to protocols which are naturally distributed. We observe that
in some realistic scenarios, speech acts in a protocol may have an eect on other
protocols. Certain actions bring about changes in the normative positions of
agents { their \social burden": what each agent is permitted, obliged and for-
bidden to do. We use the term normative command to refer to the addition or
removal of a normative position. Henceforth we shall refer to the application of
a normative command as the addition or removal of a given normative position.
Occurrences of normative positions in one protocol may also have consequences
for other protocols.
We propose to extend the notion of MAS, regulated by protocols, with an
extra layer called normative structure (NS). This layer consists of normative
scenes, which represent the normative state, i.e. the set of illocutions uttered
and normative positions, of the agents participating in a given activity, and
normative transitions, which species by means of a rule the conditions under
which some normative positions are to be generated or removed in the given
normative scenes. The formal denition of normative structure is presented in
[7], and here we informally discuss it.
Fig. 2 shows an example of how a normative structure relates with the coor-
dination level. A normative transition is specied between the negotiation and
payment activities denoting that there is a rule that may be activated with the
state of negotiation activity and that may modify the state of the payment ac-
tivity. In our example, the rule would be that whenever a client accepts an oer
of a supplier, an obligation on the former to pay the latter is created in the
payment activity. The rule connecting the payment and the work activity would

specify that whenever a client fulls its payment obligation, an obligation on the
worker to complete the contracted task is generated in the work activity.
We are concerned with the propagation and distribution of normative po-
sitions within a network of distributed, normative scenes as a consequence of
agents' actions. In [7] the
formal semantics of NSs
was dened via a map-
ping to Coloured Petri Nets.
Con
icts may arise after
the addition of new for-
mulae. Hence, if a new
norm does not generate
any con
ict then it can be
directly added. If a con-
ict arises, the algorithm
presented in [11] is used
to decide whether to ig-
nore the new normative
position or to remove the
Payment
Work
Normative Level
Exit
Registration
Payment
Work
Negotiation
Coordination Level
nt
Negotiation
Fig. 2: Normative Structure and Coordination Level
con
icting ones.
4 Proposed Distributed Architecture
We propose an architecture to address the regulation of the behaviour of au-
tonomous agents and the management of the normative state(s) of the MASs,
including the propagation of normative positions and the resolution of normative
con
icts. We assume the existence of a set of agents that interact in order to
pursue their goals { we do not have control on these agents' internal functioning,
nor can we anticipate it. We require the following features of our architecture:
Regulated The main goal of our architecture is to restrict the eects of agent
behaviour in the specied conditions without hindering the autonomy of
external agents.
Open Instead of reprogramming the MAS for each set of external agents, we ad-
vocate persistent, longer-lasting MASs where agents can join and leave them.
However, agents' movements may be restricted in certain circumstances.
Heterogeneous We leave to each agent programmer the decision of which agent
architecture include in each external agent. We make no assumption concern-
ing how agents are implemented.
Mediatory As we do not control external agents internal functioning, in order
to avoid undesired or unanticipated interactions, our architecture should
work as a \lter" of messages between agents.
Distributed To provide the means for implementing large regulated MAS, we
require our architecture to be distributed in a network and therefore spread-
ing and alleviating the workload and the message trac.

Norm propagative Although being distributed, agent interactions are not iso-
lated and agent behaviour may have eects, in the form of addition or re-
moval of normative positions, in later interactions possibly involving dierent
agents.
Con
ict Resolutive Some con
icts may arise due to normative positions being
generated as result of agent's behaviour. Since ensuring a con
ict-free MAS
at design time is computationally intractable, we require that resolution of
normative con
icts would be applied by the MAS. This approach promotes
consistency since there is a unique, valid normative state established by the
system instead of a lot of dierent state versions due to a con
ict resolution
at agent's level.
To accomplish these requirements, we extend AMELI, the architecture pre-
sented in [3]. That architecture is divided in three layers:
Autonomous agent layer It is formed by the set of external agents taking
part in the MAS.
Social layer An infrastructure that mediates and facilitates agents' interac-
tions while enforcing MAS rules.
Communication layer In charge of providing a reliable and orderly transport
service.
External agents intending to communicate with other external agents need to
redirect their messages through the social layer which is in charge of forwarding
the messages (attempts of communication) to the communication layer. In spec-
ied conditions, erroneous or il-
licit messages may be ignored by
the social layer in order to pre-
vent them from arriving at their
addressees.
The social layer presented in [3]
is a multi-agent system itself
and the agents belonging to it
are called internal agents. We
propose to extend this architec-
ture by including a new type of
agent , the normative manager
(NM1 to NMp in g. 3), and
by adding protocols to accom-
modate this kind of agent. We
Autonomous
Agents
Layer
Communication Layer
. . .
. . .
. . . . . .
. . .. . .
Distributed,
Social
Layer
P
r
i
v
a
t
e
P
u
b
l
i
c
A 1 A i A n
G 1 G i G n
SM1 SMm
N M1 N M p
Fig. 3: AMELI+architecture
call AMELI+ the resulting architecture.
In AMELI+, internal (administrative) agents are of one of the following types:
Governor (G) Internal agent representing an external agent, that is, maintain-
ing and informing about its social state, deciding or forwarding whether an
attempt from its external agent is valid. One per external agent.

Scene Manager (SM) Internal agent maintaining the state of the activity1,
deciding whether an attempt to communicate is valid, notifying any changes
to normative managers and resolving con
icts.
Normative Manager (NM) This new type of internal agent receives norma-
tive commands and may re one or more normative transition rules.
In principle, only one NM is needed if it manages all the normative transition
rules. However, in order to build large MAS and avoid bottlenecks, we propose
the distribution of rules into several NMs.
To choose the granularity of the normative layer, i.e. to choose from one
single NM to one NM per normative transition, is an important design de-
cision that we leave for
the MAS designers. Af-
ter choosing the granu-
larity, the NMs are as-
signed to handle a pos-
sibly unary set of nor-
mative transitions. Recall
that each normative tran-
sition includes a rule. The
SMs involved in the ring
of the rules are given a
reference to the NM that
manages the rule, i.e. its
NMi
Fig. 4: Channels involved in the activation of a rule
address or identier depending on the communication layer. External agents
may join and leave activities, always following the conventions of the activities.
In these cases, its governor registers (or deregisters) with the SM of that scene.
4.1 Social Layer Protocols
Fig. 4 shows the communication within the social layer { it only occurs along
the following types of channels:
Agent / Governor This type of channel is used by the external agents sending
messages to their respective governors to request information or to request
a message to be delivered to another external agent (following the norms of
the MAS). Governors use this type of channel to inform their agents about
new normative positions generated.
Governor / Scene Manager Governors use this type of channel to propagate
unresolved attempts to communicate or normative commands generated as
a result of such attempts. SMs use this type of channel to inform governors
in their scenes about new normative commands generated as a result of
attempts to communicate or con
ict resolution.
1 Hereafter, activities are also referred to as scenes following the nomenclature of
AMELI.

Scene Manager / Normative Manager This type of channel is used by SMs
to propagate normative commands that NMs may need to receive and the
ones resulting from con
ict resolution. NMs use this channel to send norma-
tive commands generated by the application of normative transition rules.
Fig. 5 shows an enactment of a MAS in our architecture. Agents send
attempts to governors (messages 1, 4 and 7) who, after nding out the
normative commands
attempts generate, prop-
agate the new norma-
tive commands to SMs1
and SMs2 (messages
2, 5 and 8) who, in
turn, propagate them
to the NM (messages
3, 6 and 9). As a nor-
mative transition rule
is red in the NM, a
NMjNM i
S Ms 1 S Ms 2
GAnne GAnneGB il l GB il l
Anne AnneB ill B ill
S Ms 3
Gpainter1 Gpaintern
painternpainter1
Fig. 5: Enactment of a normative transition rule
normative command is sent to SMs3 (message 10). After resolving any con
icts,
SMs3 sends the new normative commands to all the involved governors (mes-
sages 11 and 110) who, in turn, send them to their represented agents (messages
12 and 120).
As the gure of the previous example shows, our architecture propagates
attempts to communicate (and their eects) from agents (shown on the bottom
of Fig 5) to the NMs (shown at the top of the gure). NMs receive events from
several SMs whose managed state may be arbitrarily large. Since NMs only need
the normative commands that may cause any of its rules to re, NMs subscribe
only to the type of normative commands they are supposed to monitor. For
instance, if a rule needs to check whether there exists a prohibition to paint in
a scene work1 and whether there exists the obligation of informing about the
completion of the painting job, then the NM will subscribe to all the normative
commands adding or removing prohibitions to paint in scene work1 as well as
all normative commands managing obligations to inform about the completion
of the painting job.
In the following algorithms,
refers to essential information for the execution
of the MAS, i.e. a portion of the state of aairs of the MAS that each internal
agent is managing. As introduced above, depending on the type of the internal
agent, it manages a dierent portion of the state of aairs of the MAS, e.g. a
governor keeps the social state of the agent, and a scene manager keeps the state
of a given scene. These algorithms dene the behaviour of internal agents and
are applied whenever a message msg is sent by an agent (agi), a governor (gi),
a SM (smi) or a NM (nmi) respectively.
When an external agent sends to its governor an attempt to communicate
(messages 1, 4 and 7 in Fig. 5), the governor follows the algorithm of Fig. 6(a).
This algorithm checks whether the attempt to communicate generates normative

algorithm G process att(agi;msg)
input agi;msg
output ;
begin
01 new cmmds := get norm cmmds(msg;
)
02 foreach c 2 new cmmds do
03
:= apply(c;
)
04 sm := scene manager(c)
05 send(c; agi)
06 send(c; sm)
07 endforeach
08 if new cmmds = ; then
09 sm := scene manager(msg)
10 send(msg; sm)
11 endif
end
(a) G response to an agent attempt
algorithm NM process cmmd(smi;msg)
input smi;msg
output ;
begin
01 foreach cmmd 2 msg do
02
:= apply(cmmd;
)
03 ncs := get RHS from fired rules(
)
04 foreach c 2 ncs do
05 sm := scene manager(c)
06 send(c; sm)
07 endforeach
08 foreach
end
(b) NM response to a command
algorithm SM process att(gi;msg)
input gi;msg
output ;
begin
01 new cmmds := get norm cmmds(msg;
)
02 foreach c 2 new cmmds do
03
:= apply(c;
)
04 send(c; gi)
05 foreach hnm; evi 2 subscriptions do
06 if unify(c; ev; ) then
07 send(c; nm)
08 endif
09 endforeach
10 endforeach
11 if new cmmds = ; then
12 s := scene(msg)
13 c := content(msg)
14 send(rejected(s; c); gi)
15 endif
end
(c) SM response to a forwarded attempt
algorithm SM process cmmd(nmi;msg)
input nmi;msg
output ;
begin
01
0 := apply(msg;
)
02 if inconsistent(
0) then
03 msg := resolve conflicts(
;msg)
04 endif
05 foreach cmmd 2 msg do
06
:= apply(cmmd;
)
07 foreach hnm; evi 2 subscriptions do
08 if unify(c; ev; ) then
09 send(c; nm)
10 endif
11 endforeach
12 foreach g 2 governors(cmmd) do
13 send(cmmd; g)
14 endforeach
15 endforeach
end
(d) SM response to a command
Fig. 6. Internal Agents Algorithms
commands (line 1), i.e. it is accepted2. This check may vary depending on the
type of specication and implementation of the scenes: e.g. using Finite State
Machines (FSM), as in [3], or executing a set of rules, as in [9].
If the attempt generates normative commands (line 2), they are applied to
the portion of the state of aairs the governor is currently managing creating a
new partial state (line 3). These normative commands are sent to the external
agent (line 5) and to the scene manager (messages 2, 5 and 8 in Fig. 5) in charge
of the scene where the normative command should be applied (line 6). Otherwise,
the attempt is forwarded to the SM of the scene the attempt was generated in
(line 10).
If the governor accepts the attempt (after the check of line 1), it sends the
SM a notication.The SM then applies the normative command received and
forwards it to the NMs subscribed to that event (messages 3, 6 and 9 in Fig. 5).
However, if the governor does not take a decision, i.e. normative commands
are not generated, the governor sends the attempt to the SM who should decide
whether it is valid or not by following the algorithm of Fig. 6(c). This algorithm,
2 In our approach, an ignored attempt would not generate any normative command.

like the one in Fig. 6(a), checks whether the received attempt generates norma-
tive commands in the current scene state, i.e. the portion of the state of aairs
referring to that scene (line 1). If this is the case (line 2), they are applied to the
current state of the scene (line 3) and forwarded to the governor that sent the
attempt (line 4) and to the NMs subscribed to that normative commands (line
7). Otherwise (line 11), a message informing that the attempt has been rejected
is sent to the governor mentioned (line 14).
In both cases, if the attempt is accepted then the normative manager is noti-
ed and it follows the algorithm of Fig. 6(b) in order to decide if it is necessary
to send new normative commands to other scene managers. This algorithm pro-
cesses each normative command received (line 1) by applying it to the state of
the NM (line 2) and checking which normative transition rules are red and
obtaining the normative commands generated (line 3). Each of them are prop-
agated to the SM of the scene appearing in the normative command (line 6,
message 10 in Fig. 5).
If normative commands are generated, SMs receive them from the normative
manager in order to resolve possible con
icts and propagate them to the appro-
priate governors. In this case, the SMs execute the algorithm of Fig. 6(d). This
algorithm applies the normative command received on the scene state creating a
temporary state for con
ict checking (line 1), then checks if the new normative
command would raise an inconsistency (line 2). If this is the case, it applies
the con
ict resolution algorithm presented in [7], returning the set of norma-
tive commands needed to resolve the con
ict (line 3). Each normative command
caused by the message sent by the NM or by con
ict resolution, is applied to
the scene state (line 6) and it is sent to the subscribed NMs (lines 7-11) and
to the governors (messages 11 and 11' in Fig. 5) of the agents appearing in the
normative command (lines 12-14).
NMs are notied about the resolution of possible con
icts in order to check
if the new normative commands re normative transition rules. If NMs receive
this notication, they follow again the algorithm of Fig. 6(b) as explained above.
When governors are notied by a SM about new normative commands, they
apply the normative command received to the normative state of the agent and
notify to its agent about the new normative command (messages 12 and 12' in
Fig. 5).
In our approach, con
ict resolution is applied at the SM level requiring all
normative commands generated by a NM to pass through a SM who resolves
con
icts and routes them. This feature is justied because SMs are the only
agents who have a full representation of a scene and know the agents are partici-
pating in it and which role they are enacting. For example, if a prohibition for all
painters to paint arrives at the work activity, a SM will forward this prohibition
to the governors of the agents participating in that activity with the painter role
and to the governors of all the new painters that join that activity while the
prohibition is active. An alternative approach is to apply con
ict resolution at
the level of governor agents, curtailing some of the normative positions of its as-
sociated external agent. However, this type of con
ict resolution is more limited

since a governor only maintains the normative state of an agent. For example,
a case that cannot be resolved with this approach is when all agents enacting a
role are simultaneously prohibited and obliged to do something, i.e. when more
than one agent is involved in the con
ict.
Another approach would be if governors became the only managers of norma-
tive positions; in this case they would need to be aware of all normative positions
that may aect its agent in the future, i.e. they would have to maintain all the
normative positions aecting any of the roles that its agent may enact in every
existing scene. For instance, a governor of an agent that is not yet enacting a
painter role would also need to receive the normative positions that now applies
to that role even if the agent is not in that scene or is enacting that role yet. This
approach does not help with scalability since a large MAS with various scenes
may generate a very large quantity of normative positions aecting agents in the
future by the mere fact of their entering the MAS.
5 Related Work
The subject of norms has been studied widely in the literature (e.g., [18,16,15]),
and, more recently, much attention is being paid to more pragmatic and imple-
mentational aspects of norms, that is, how norms can be given a computational
interpretation and how norms can be factored in the design and execution of
MASs (e.g. [1,2,5,9,8]).
However, not much work has addressed the management of norms and rea-
soning about them in a distributed manner. Despite the fact that in [4,12] two
languages are presented for the distributed enforcement of norms in MAS, in
both works each agent has a local message interface that forwards legal mes-
sages according to a set of norms. Since these interfaces are local to each agent,
norms can only be expressed in terms of actions of that agent. This is a serious
disadvantage, e.g. when one needs to activate an obligation to one agent due to
a certain message of another agent.
In [17] the authors propose a multi-agent architecture for policy monitoring,
compliance checking and enforcement in virtual organisations (VOs). Their ap-
proach also uses a notion of hierarchical enforcement, i.e. the parent assimilates
summarised event streams from multiple agents and may initiate further ac-
tion on the subordinate agents. Depending on its policies, a parent can override
the functioning of its children by changing their policies. Instead of consider-
ing any notion similar to our scene (multi-agent protocol where the number of
participants may vary) and assigning an agent exclusively dedicated to the man-
agement of one scene, they assign another participant in the VO as parent of a
set of agents. Although the parent would receive only the events it needs to mon-
itor, it may receive them from all the interactions their children are engaging in.
This can be a disadvantage when the number of interactions is large converting
the parents in bottlenecks. Although they mention that con
ict resolution may
be accomplished with their architecture, they leave this feature to the VO agent
thus centralising the con
ict resolution in each VO. This can also be a disadvan-

tage when the number of interactions is large since the VO agent has to resolve
all the possible con
icts. This would require either all the events
owing through
the VO agent or the VO agent monitoring the state of the whole VO in order to
detect and resolve con
icts. The main theoretical restriction in their approach is
that all the agents involved in a change in a policy must share a common parent
in the hierarchy of the VO. In an e-commerce example, when a buyer accepts a
deal an obligation to supply the purchased item should be added to the seller.
However, as they are dierent parties, their only common parent is the VO agent
converting the latter in a bottleneck in large e-commerce scenarios.
6 Conclusions and Future Work
We base the architecture presented in this paper in our proposal of normative
structure and con
ict resolution of [7]. The notion of normative structure is
useful because it allows the separation of normative and procedural concerns.
We notice that the algorithm presented in that paper is also amenable to the
resolution of normative con
icts in a distributed manner.
The main contribution of this paper is an architecture for the management
of norms in a distributed manner. As a result of the partial enactment of pro-
tocols in diverse scenes, normative positions generated in dierent scenes can
be used to regulate the behaviour of agents not directly involved in previous
interactions. Furthermore, con
ict resolution is applied at a scene level meaning
that resolution criteria involving more than one agent are now possible.
We want to extend normative structures [7], as we use them in our archi-
tecture, along several directions: (1) to handle constraints as part of the norm
language, in particular constraints related with the notion of time; (2) to capture
in the con
ict resolution algorithm dierent semantics relating the deontic no-
tions by supporting dierent axiomations (e.g., relative strength of prohibition
versus obligation, default deontic notions, deontic inconsistencies, etc.).
We also intend to use analysis techniques for Coloured Petri-Nets (CPNs)
in order to characterise classes of CPNs (e.g., acyclic, symmetric, etc.) corre-
sponding to families of Normative Structures that are susceptible to tractable
o-line con
ict detection. The combination of these techniques along with our
online con
ict resolution mechanisms is intended to endow MAS designers with
the ability to incorporate norms into their systems in a principled way.
Acknowledgements { This work was partially funded by the Spanish Edu-
cation and Science Ministry as part of the projects TIN2006-15662-C02-01 and
2006-5-0I-099. Garca-Camino enjoys an I3P grant from the Spanish National
Research Council (CSIC).
References
1. A. Artikis, L. Kamara, J. Pitt, and M. Sergot. A Protocol for Resource Shar-
ing in Norm-Governed Ad Hoc Networks. In Declarative Agent Languages and
Technologies II, volume 3476 (LNCS). Springer-Verlag, 2005.

2. S. Craneeld. A Rule Language for Modelling and Monitoring Social Expectations
in Multi-Agent Systems. Technical Report 2005/01, University of Otago, 2005.
3. M. Esteva, B. Rosell, J. A. Rodrguez-Aguilar, and J. L. Arcos. AMELI: An
agent-based middleware for electronic institutions. In Procs of 3rd Int'l Conf on
Autonomous Agents and Multiagent Systems (AAMAS'04), pages 236{243, 2004.
4. M. Esteva, W. Vasconcelos, C. Sierra, and J. A. Rodrguez-Aguilar. Norm consis-
tency in electronic institutions. In XVII Brazilian Symposium on Articial Intel-
ligence - SBIA'04, volume 3171 (LNAI), pages 494{505. Springer-Verlag, 2004.
5. N. Fornara, F. Vigano, and M. Colombetti. An Event Driven Approach to Norms
in Articial Institutions. In AAMAS05 Workshop: Agents, Norms and Institutions
for Regulated Multiagent Systems (ANI@REM), Utrecht, 2005.
6. Foundation for Intelligent Physical Agents (FIPA). FIPA-ACL: Message Structure
Specication, December 2002.
7. D. Gaertner, A. Garca-Camino, P. Noriega, J. A. Rodrguez-Aguilar, and W. Vas-
concelos. Distributed Norm Management in Regulated Multi-agent Systems. In
Procs of 6th Int'l Conf on Autonomous Agents and Multiagent Systems (AA-
MAS'07), pages 624{631, Hawai'i, May 2007.
8. A. Garca-Camino, P. Noriega, and J. A. Rodrguez-Aguilar. Implementing Norms
in Electronic Institutions. In Procs of 4th Int'l Conf on Autonomous Agents and
Multiagent Systems (AAMAS'05), pages 667{673, Utrecht, July 2005.
9. A. Garca-Camino, J.-A. Rodrguez-Aguilar, C. Sierra, and W. Vasconcelos. A Dis-
tributed Architecture for Norm-Aware Agent Societies. In Decl. Agent Languages
and Technologies III, volume 3904 (LNAI), pages 89{105. Springer, 2006.
10. J. Habermas. The Theory of Communication Action, Volume One, Reason and
the Rationalization of Society. Beacon Press, 1984.
11. M. J. Kollingbaum, W. W. Vasconcelos, A. Garca-Camino, and T. J. Norman.
Con
ict resolution in norm-regulated environments via uni cation and constraints.
In Fifth International Workshop on Declarative Agent Languages and Technologies
(DALT 2007), Hawai'i, May 2007.
12. N. Minsky. Law Governed Interaction (LGI): A Distributed Coordination and
Control Mechanism (An Introduction, and a Reference Manual). Technical report,
Rutgers University, 2005.
13. A. Ricci and M. Viroli. Coordination Artifacts: A Unifying Abstraction for En-
gineering Environment-Mediated Coordination in MAS. Informatica, 29:433{443,
2005.
14. J. Searle. Speech Acts, An Essay in the Philosophy of Language. Cambridge Uni-
versity Press, 1969.
15. M. Sergot. A Computational Theory of Normative Positions. ACM Trans. Comput.
Logic, 2(4):581{622, 2001.
16. Y. Shoham and M. Tennenholtz. On Social Laws for Articial Agent Societies:
O-line Design. Articial Intelligence, 73(1-2):231{252, 1995.
17. Y. B. Udupi and M. P. Singh. Multiagent policy architecture for virtual bussiness
organizations. In Proceedings of the IEEE International Conference on Services
Computing (SCC), September 2006.
18. G. H. von Wright. Norm and Action: A Logical Inquiry. Routledge and Kegan
Paul, London, 1963.