A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent

Systems

essica Soares dos Santos

and Viviane Torres da Silva

Computer Science Department, Universidade Federal Fluminense, Niter

oi, Brazil

IBM Research (on leave from Universidade Federal Fluminense), Rio de Janeiro, Brazil

Keywords:

Multi-agent Systems, Norms, Conﬂict Detection, Ontology, WordNet.

Abstract:

Norms are usually applied in Multi-Agent Systems to regulate the behavior of software agents and maintain

social order. Those systems can be regulated by multiple norms and require a mechanism to verify whether

the set of norms is conﬂict-free or not. The detection of indirect normative conﬂicts is not a trivial task since

they only can be identiﬁed when the detection mechanism is able to infer that different elements that compose

two norms are related in some way. In this research, we propose a mechanism to detect normative conﬂicts

by combining two different approaches. The former uses information from a domain ontology that stores

relationships that are exclusive of the MAS. The latter uses information from a lexical database called WordNet

that stores relationships among concepts of the real world. This research results in the implementation of a

tool with a robust mechanism for normative conﬂict detection that can be used during the design of a MAS.

1 INTRODUCTION

In Multi-agent Systems (MAS), norms are being used

in order to control and restrict the software agents be-

havior. A way of regulation is needed in those sy-

stems to avoid the occurrence of undesirable actions

since software agents are autonomous entities that can

be independently designed. When MAS are gover-

ned by multiple norms, it is essential the existence

of a mechanism to verify whether such norms contra-

dict each other or not. When there is a contradiction

between two norms addressed to the same agent of

a MAS we say that there is a normative conﬂict be-

tween the norms. In such a case, the agent cannot

comply with both norms simultaneously without vio-

lating one of them. There are two kinds of normative

conﬂicts known in the literature, as follows: (i) di-

rect conﬂicts: they are conﬂicts that involve norms ad-

dressed to the same elements but that have contradic-

tory or opposite modalities, i.e., a prohibition versus

an obligation or permission regulating the same beha-

vior; and (ii) indirect conﬂicts: they are conﬂicts that

involve norms that are addressed to different but rela-

ted elements. Direct conﬂicts can be easily detected

through a direct analysis of the norm elements. On the

other hand, it is a challenge to detect indirect conﬂicts

since the conﬂicting norms are addressed to different

elements and the conﬂict can only be detected when

relationships among the norm elements are identiﬁed.

The detection of indirect conﬂicts among norms is a

topic that is being widely studied in MAS. However,

the detection processes of all approaches that we have

surveyed in the literature are only possible when the

application designer speciﬁes relationships among the

elements that compose the norms of the system in a

document or an axiom, for instance (see Section 2). A

way of detecting indirect conﬂicts by using the lexi-

cal database WordNet (Miller, 1995) was described in

(Santos and Silva, 2016). The WordNet stores words

and semantic relationships among them, for instance,

the words “start” and “stop” are verbs that are related

in WordNet because they denote opposite actions in

the real world. In this example, such a relationship

can be used to detect indirect conﬂicts between two

norms where one obliges the agent to start and the

other obliges the same agent to stop. Using the Word-

Net, the application designer does not need to worry

about specifying relationships that are not speciﬁc of

the MAS (domain-independent relationships). Addi-

tionally, another advantage is that norms of a MAS

can be deﬁned by different designers and WordNet

can be useful to unify the syntax of the norms, as syn-

onymous words are stored together in WordNet. Ho-

wever, a MAS may need to consider speciﬁc relati-

onships, i.e., relationships that do not exist in the real

world but exist in the MAS. For instance, relations-

Santos, J. and Silva, V.

A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent Systems.

DOI: 10.5220/0006598700700079

In Proceedings of the 10th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2018) - Volume 1, pages 70-79

ISBN: 978-989-758-275-2

hips that specify that a given software agent inhabits

a given environment of the MAS. Those relationships

cannot be captured only by using the WordNet. For

this reason, in this paper, we present a novel mecha-

nism that combines an approach of conﬂict detection

that uses the WordNet with an approach that uses an

ontology to identify domain dependent relationships

and detect indirect conﬂicts among norms of a MAS.

By using this mechanism, the application designer

only needs to specify in a domain ontology relati-

onships that are exclusive of the given MAS. The re-

mainder of this paper is organized as follows: Section

2 presents other approaches that deal with conﬂicts

among norms in MAS. Section 3 presents all back-

ground information needed to the understanding of

our research. Section 4 describes our tool and the

steps performed by the conﬂict detection mechanism.

Section 5 presents a case study in order to demon-

strate the execution of our mechanism. In Section 6,

we present our conclusions, highlight limitations and

point out suggestions for future work.

2 RELATED WORK

There are many approaches in the literature that pro-

pose a means to detect conﬂicts among norms of a

MAS. Most of them can detect direct conﬂicts (Neto;

Silva; Lucena, 2012), (Neto; Silva; Lucena, 2013),

(Gaertner et al., 2007), (Vasconcelos et al., 2012)

and others can also detect some kinds of indirect

conﬂicts (S¸ensoy, et al., 2012), (Aphale; Norman;

S¸ensoy, 2013), (Kollingbaum et al., 2007), (Vascon-

celos; Kollingbaum; Norman, 2009), (Zahn, 2015),

(Fenech; Pace; Schneider, 2008), (Giannikis and Das-

kalopulu, 2011). The strategies of detection of con-

ﬂicts can be classiﬁed according to when the strategy

is applied. The detection process can occur at de-

sign time or at runtime. Design time strategies (Fe-

nech; Pace; Schneider, 2008), (S¸ ensoy, et al., 2012),

(Aphale; Norman; S¸ ensoy, 2013), (Zahn, 2015) de-

tect conﬂicts during the design/speciﬁcation phase of

the MAS, that is, before the execution of the MAS.

In this case, normative conﬂicts are detected before

they occur and the system designer can verify if the

set of norms of a MAS is conﬂict-free. A disadvan-

tage of adopting design time strategies is that some

conﬂicts only can be detected during the execution of

the system (conﬂicts that depend on the execution or-

der of runtime events). On the other hand, runtime

strategies (Kollingbaum et al., 2007), (Vasconcelos;

Kollingbaum; Norman, 2009), (Giannikis and Daska-

lopulu, 2011) detect conﬂicts during the execution of

the MAS. Usually, when an approach uses a runtime

detection strategy, it also presents a strategy to resolve

conﬂicts. A disadvantage of runtime strategies is that

they can be computationally expensive and it can im-

pact on the performance of the MAS.

Most approaches that can detect indirect norma-

tive conﬂicts consider relationships among actions

to do their analysis. The researches presented in

(S¸ensoy, et al., 2012), (Aphale; Norman; S¸ ensoy,

2013), (Kollingbaum et al., 2007) and in (Vasconce-

los; Kollingbaum; Norman, 2009) analyze the side-

effects of the performance of the actions in order to

detect indirect conﬂicts. The work in (Vasconcelos;

Kollingbaum; Norman, 2009) and the approach des-

cribed in (Zahn, 2015) verify if actions are related by

a composition relationship in order to detect conﬂicts.

The researches in (Fenech; Pace; Schneider, 2008),

(Giannikis and Daskalopulu, 2011) and (Zahn, 2015)

take into account a relationship of orthogonality bet-

ween actions, which relate actions that cannot be per-

formed at same time. The work in (Zahn, 2015) also

considers hierarchy among entities, and relationships

that relate an entity to the environment that it inhabits,

or an entity to the role it plays, for instance.

Our research differs from the others because it can

detect indirect conﬂicts that occur due to relations-

hips that were not deﬁned by the application designer,

and in addition of considering all relationships among

actions, entities and contexts deﬁned by (Zahn, 2015),

we also consider other relationships that are not con-

sidered by other approaches, such as, synonymy and

antonymy as detailed in Section 3.

We present a deeper discussion about techniques

of detection and resolution of normative conﬂicts in

MAS in a previous work (Santos et al., 2017).

3 BACKGROUND

In this section, we present the essential aspects of our

proposal. First, we present the norm deﬁnition adop-

ted. The norm deﬁnition is an important factor, since

all kinds of conﬂicts that can be detected depend on

the norm expressivity, i.e., the elements that a norm

can represent. After that, we list the relationships that

the mechanism to detect conﬂicts will investigate be-

tween the elements of the norms. Our research com-

bines two different approaches of conﬂict detection

whose relationships are listed separately.

3.1 Norm Deﬁnition

We assume that a norm obliges, permits or prohibits

an entity to perform an action that can be applied to

a speciﬁc object. For instance, a norm can say that

A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent Systems

an agent is obliged to drive a car, where drive is the

action and car is the object. The object is an optional

element of our norm deﬁnition, i.e., a norm can regu-

late an action that is not associated with an object.

Deﬁnition: A norm is a tuple in the form

n = hid, deoC, c, e, act(ob j), ac, dci

where id is the norm identiﬁer; deoC is the deon-

tic concept that determines the modality of the norm

deoC ∈ {obligation, permission, prohibition}; c ∈ C

is the context where the norm is deﬁned (it can be an

organization o ∈ O or an environment env ∈ Env); e ∈

E is the entity being regulated by the norm. An entity

e may be an agent a ∈ A, an organization o ∈ Org or

a role r ∈ R; act ∈ Act is the action being regulated;

obj ∈ Obj is the object associated with the action. The

object is an optional ﬁeld; and ac ∈ Cd and dc ∈ Cd

are dates that, respectively, activate and deactivate the

norm. The symbol “ ” can be used to determine that

a norm regulates all entities of a speciﬁc context.

3.2 Domain Ontology Relationships

The mechanism proposed to detect conﬂicts is able to

receive an ontology as input, specifying the relations-

hips of the application domain. The relationships that

can be speciﬁed in the ontology by the application de-

signer are listed below, as follows:

Inhabit: it relates an entity to the environment

that it inhabits. This relationship indicates that if a

norm regulates an environment, the norm also regula-

tes the entities that inhabit such an environment.

Play: it relates an entity to the roles it can assume.

This relationship indicates that if a norm regulates a

role, the norm also regulates the entities that play such

a role.

Ownership: it deﬁnes the roles that belong to a

given organization. This relationship indicates that if

a norm regulates an organization, the norm also regu-

lates the roles that belong to the organization.

Hierarchy: it deﬁnes that an element is super ele-

ment of another one. This relationship indicates that

if there is a norm regulating a super context/entity, the

norm also regulates their sub contexts/entities.

Reﬁnement: it deﬁnes that an action is the speci-

alization of another one.

Composition: it deﬁnes that an action (called

whole action) is composed of other actions (called

part actions).

Orthogonality: it deﬁnes actions that cannot be

performed at the same time by the same entity or re-

lated entities.

Dependency: it determines that an action (called

client action) is a precondition to the performance of

another one (called dependent action).

3.3 WordNet Relationships

Our approach uses the WordNet database to ﬁnd re-

lationships between contexts, entities, actions and ob-

jects (associated with actions).

Synonymy: words that denote the same concept

are grouped in a same set (called synset) in the Word-

Net and are related by the relationship Synonymy. We

map the contexts described in the norms to nouns and

verify if they are related through the relationship Syn-

onymy. For instance, if there is a norm whose context

is United States of America and other one whose con-

text is USA, the algorithm will conclude that both con-

texts are equivalents and that both norms are applied

to the same context. Similarly, by using the relations-

hip Synonymy we can infer that two norms that are, in

principle, addressed to different entities, in fact, refer

to the same entity. For instance, physician and doc-

tor are nouns that denote a licensed medical practiti-

oner. We also use the relationship Synonymy to map

the actions to verbs and objects to nouns, and identify

that the actions/objects of two norms are equivalent.

For instance, the actions to collaborate and to coope-

rate are related by the relationship Synonymy.

Specialization: it is described in the WordNet as

“Hyponymy/Hypernymy” and relate a noun that deno-

tes an element to its respective sub-elements contexts

and super-contexts. For instance, the context hospital

is a sub context of medical institution, since a hospi-

tal is a medical institution. Similarly, this relationship

can be used to detect that a sub-entity is related to a

super-entity. When a norm is applied to a super-entity,

such a norm is propagated to its sub-entities. For in-

stance, if a norm is applied to the role doctor, suppo-

sing that there is the role angiologist in the domain,

the norm also is applied to all entities that are playing

the role angiologist. We also use this relationship to

detect relationships among objects. For instance, the

object train is a specialization of the object public

transport, because a train is a kind of public trans-

port. The relationship Specialization among verbs is

deﬁned as “Troponymy/Hypernymy” in the WordNet

and is used to identify that a sub-action is related to a

super-action. For instance, in the WordNet the verbs

to move and to walk are related by this relationship

since to walk is a way of to move.

Part-Whole: it is described in the WordNet as

“Meronymy/Holonymy” and relates an element that

denotes a part to an element that denotes a whole.

Usually, this relationship is used to relate geographic

areas in the WordNet. For instance, the context USA

is part of the context North America. Similarly, the

intensive care unit is part of the hospital and sacristy

is a part of church. This relationship can be app-

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

lied to detect related objects, for instance, the objects

window and car are related by the relationship Part-

Whole because a window is part of a car.

Entailment: it relates an action that entails anot-

her one. For instance, when someone buy something

it must pay for it. Then, the verbs to buy and to pay

are related in the WordNet by the relationship Entail-

ment.

Antonymy: it relates an action to its opposite. For

instance, to move and to stop are verbs related in the

WordNet by the relationship Antonymy because they

denote opposite actions.

4 CONFLICT CHECKER TOOL

In this section, we describe how our proposal of con-

ﬂict checker

was implemented. Our tool was deve-

loped using Java programming language and NetBe-

ans IDE 8.0 integrated with the OWL-API (Horridge

and Bechhofer, 2011), which is the API that we use

to read an OWL ontology. The tool uses the JWNL

library (Walenz and Didion, 2011) to perform an off-

line searching in the WordNet. Our algorithm recei-

ves as input a domain ontology that describes dom-

ain elements (contexts, entities, actions, objects), the

set of norms and, optionally, domain-dependent rela-

tionships. After that, it combines two approaches of

conﬂict detection and is divided into two steps, as fol-

lows:

1. Detection of domain-dependent conﬂicts by using

relationships described in the domain ontology

(Domain Conﬂict Checker);

2. Detection of domain-independent conﬂicts by

using relationships described in the WordNet

(WordNet Conﬂict Checker).

Both approaches are divided into the following

sub-steps:

a. Propagation of norms according to the relations-

hips;

b. Grouping norms in sets according to their simila-

rity;

c. Veriﬁcation of time intersection between each pair

of norms belonging to the same set;

d. Application of rules to identify conﬂict patterns;

After receiving the domain ontology describing

the relationships, entities, contexts, actions, objects

and the norms considered by the MAS, the algorithm

performs a propagation of norms. During propaga-

tion, norms addressed to general entities and contexts

goo.gl/7TmYPX

Figure 1: Example of norm propagation.

are addressed to speciﬁc entities and contexts of the

domain (sub-step a). For instance, let us consider that

there are two agents called agent1 and agent2 that are

playing the role physician in the MAS (relationship

play). Then, if the application designer speciﬁes a

norm n1 that is addressed to the role physician, the

propagation process will create two new norms n1.1

and n1.2 that will be composed of the same elements

of n1 but will be addressed to the agents agent1 and

agent2 (see Figure 1). However, the propagation of

norms according to domain-dependent relationships

can generate inconsistent norms if it contradicts other

relationships deﬁned in the domain ontology. When

it occurs, such norms must be discarded. The rules to

discard inconsistent norms are described in Table 1.

Since the propagation of contexts and entities can ge-

nerate multiple norms, to reduce the number of com-

parisons needed, norms that have the same entity and

context are grouped in the same group (sub-step b).

For instance, suppose that exists a set of norms: n3,

n4, n5, n6, n7 and n8. The norms n3, n6 and n8 are

associated with the context Brazil and regulate the en-

tity agent3; the norm n5 is associated with the con-

text Argentina and regulates the entity agent4; and

the norms n4 and n7 are addressed to the agent5 and

to the context USA and United States of America, re-

spectively. Note that the contexts USA and United

States of America are synonyms. Then, in this exam-

ple, the sub-step of grouping norms will create three

sets of norms (see Figure 2) and only norms belonging

to a same set will be compared in the next sub-step.

Only norms addressed to the same (or equivalent) en-

tities and applied to the same (or equivalent) contexts

Figure 2: Example of grouping norms according to their

similarity.

A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent Systems

Table 1: Rules to discard inconsistent norms generated by

the propagation process.

Kind of

propagation

Discard rule

Hierarchy

among con-

texts that are

environments

The entity of the original norm is

an organization/agent and is not de-

ﬁned in the domain ontology that

such entity inhabits the environ-

ment of the propagated norm (inha-

bit)

Hierarchy

among con-

texts that are

organizations

(i) the entity of the original norm is

an organization that is not a suborg-

anization (hierarchy) of the context

of the propagated norm and is not

the context of the propagated norm

(ii) the entity of the original norm is

a role that is not related to the or-

ganization that is the context of the

propagated norm (ownership)

Inhabit bet-

ween contexts

that are envi-

ronments and

contexts

that are orga-

nizations

(i) the entity of the original norm is

an organization that is not a suborg-

anization (hierarchy) of the context

of the propagated norm and is not

equal to the context of the propaga-

ted norm

(ii) the entity of the original norm

is a role that is not related (owners-

hip) to the context of the propagated

norm (that is an organization)

Inhabit bet-

ween contexts

that are envi-

ronments and

entities

that are orga-

nizations

The entity of the propagated norm

is not a suborganization (hierarchy)

of the entity of the original norm

can conﬂict. Next, for each pair of norms of a same

group, the algorithm veriﬁes if there is an intersection

between the activation and deactivation conditions of

the given two norms (sub-step c) and, if so, the al-

gorithm analyzes the actions, objects and the deontic

concepts of the norms in order to verify if the pair of

norms is conﬂicting (sub-step d). The norms that were

propagated and the conﬂicts detected are passed to

the second step as an input parameter. The algorithm

performs norm propagation considering relationships

described in the WordNet (sub-step a). Norms are

grouped together in sets when they are addressed to

the same or to a synonym entities and contexts (sub-

step b). The algorithm veriﬁes if there is a time in-

tersection to each pair of norms of the same set (sub-

step c). The algorithm applies the conﬂict rules to

detect conﬂicting patterns considering the relations-

hips of WordNet (sub-step d). To conclude that two

norms are in conﬂict, the actions deﬁned in the norms

must be analyzed with their objects (when the action

involves objects). The objects of norms are mapped to

WordNet nouns and their relationships are analyzed.

Finally, the algorithm exhibits all conﬂicts detected

and the domain-dependent and domain-independent

relationships that were identiﬁed. Note that in Section

5, we exhibit some conﬂicting patterns, that is, pat-

terns that indicate that two norms cannot be adopted

at the same time and for this reason are in conﬂict.

The complete list of conﬂicting patterns that have

been deﬁned involving domain-independent/domain-

dependent relationships are described in (Santos and

Silva, 2016) and (Zahn, 2015), respectively.

Figure 3 illustrates the graphical interface of the

conﬂict checker tool. The button “Upload” is used

to select an OWL ontology. The button “Execute”

performs the detection process. The ﬁeld “Onto-

logy Description” exhibits the information described

in the ontology received as input, that is, norms, con-

texts, entities, actions, objects and relationships of the

domain application described by the application de-

signer. The ﬁeld “Domain Conﬂict Checker” exhi-

bits the propagations due to domain-dependent rela-

tionships, the pairs of norms compared and indicates

whether each pair is conﬂicting or not and the rea-

son. Similarly, the ﬁeld “WordNet Conﬂict Checker”

exhibits the propagations due to domain-independent

relationships, exhibits the pairs of norms compared

and indicates the reason of conﬂict. The ﬁeld “Con-

clusion” groups information from the ﬁelds “Domain

Conﬂict Checker” and “WordNet Conﬂict Checker”

and exhibits all conﬂicting patterns found to each pair

of norms compared. The “Reset” button cleans all in-

formation exhibited on the screen to be able to select a

new ontology and perform another detection process.

Figure 3: GUI of the Conﬂict Checker.

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

5 CASE STUDY

In this section, we present a case study to illustrate the

process of conﬂict detection. Our tool will use norms

of an e-commerce scenario.

5.1 Deﬁnition of the Case Study

The type of e-commerce contract of this case study

is an agreement involving two parties called contrac-

tor and contracted, where the contracted provides a

service of management and availability of a virtual

shop and the contractor is the entity that has interest

in acquiring such a service, that is, it has an interest

in making its products available for sale on the In-

ternet. The selected contract is a default agreement

that deﬁnes the basic conditions for the establishment

of an agreement between the contracted and any con-

tractor, where both inhabit the context of USA. In the

following case study, the deontic concepts “obliga-

tion”, “permission” and “prohibition” are represented

as “O”, “P”, and “F”, respectively. Since the selected

contract has many norms, we have selected only a few

to demonstrate the process of conﬂict detection in a

simpliﬁed way, as follows:

1) The contractor is prohibited from being una-

ware of any part of the contract

hN1, F, USA, contractor, unaware(contract), , i

2) The contractor is obliged to agree to the con-

tract

hN2, O, USA, contractor, agree(contract), , i

3) The contractor cannot reproduce the tools pro-

vided by the contracted. Such tools are software pro-

grams of the virtual shop management

hN3, F, USA, contractor, reproduce(tool), , i

4) The contractor also cannot market the tools

made available by the contracted

hN4, F, USA, contractor, market(tool), , i

5) The contractor cannot send e-mails to users of

the virtual shop

hN5, F, USA, contractor, send(email), , i

6) The contractor must purchase the software to

be installed in the virtual shop environment

hN6, O, USA, contractor, purchase(so ftware), , i

7) The contractor is obliged to deliver the products

sold in the website

hN7, O, USA, contractor, deliver(product), , i

8) The contractor is prohibited to access servers of

the contracted

hN8, F, USA, contractor, access(server), , i

9) The contracted may exclude products from the

virtual shop, for instance, when it considers that the

sale of them is improper

hN9, P, USA, contracted, exclude(product), , i

10) The contracted may change the contract

hN10, P, USA, contracted, change(contract), , i

11) The contractor is obliged to include a freight

to the sales

hN11, O, USA, contractor, include( f reight), , i

In order to demonstrate the operation of our pro-

posal for the detection of conﬂicts, we will consider

that there is a contractor party called Company 1 that

has an interest in using the services of the contrac-

ted party and that also has preconditions in relation

to the aforementioned agreement with the contracted

party. Therefore, our conﬂict resolution consists in

analyzing the norms of the default contract deﬁned by

the contracted party together with the speciﬁc norms

of a given contractor in order to identify inconsisten-

ces. Let us assume that Company 1 has the following

norms in relation to the default contract established

by the contracted:

12) Company 1 may be unaware of clauses of the

contract

hN12, P, USA, Company 1, unaware(clause), , i

13) Company 1 may disagree with the contract

clauses

hN13, P, USA, Company 1, disagree(clause), , i

14) Company 1 may copy tools from the contrac-

ted

hN14, P, USA, Company 1, copy(tool), , i

15) Company 1 may login on computers from

contracted

hN15, P, USA, Company 1, login(computer), , i

16) Company 1 may exclude the freight of pur-

chases made in California during the month of De-

cember 2017

hN16, P, Cali f ornia, Company 1, exclude( f reight),

12/01/2017 00:00:00, 12/31/2017 00:00:00i

17) The contracted cannot change the contract

hN17, F, USA, contracted, change(contract), , i

18) Company 1 may send spam, for instance, to

promote the sale of products or other services

hN18, F, USA, Company 1, send(spam), , i

19) The contracted cannot perform actions associ-

ated with the management of products from the vir-

tual shop, that is, it cannot insert, edit or exclude pro-

ducts

hN19, F, USA, contracted, manage(product),

, i

20) Company 1 cannot pay for any software that

will be installed in the virtual shop

hN20, F, USA, Company 1, pay(so f tware), , i

The application designer only needs to explicitly

describe in the ontology the existing relationships of

the application domain. Thus, in this case it was only

necessary to explicitly deﬁne the following relations-

hips within the domain ontology: (i) Play is declared

A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent Systems

between Company 1 and contractor, indicating that

Company 1 is an entity that plays the role of contrac-

tor; and (ii) Reﬁnement is declared between the action

manage and the actions register, edit and delete.

5.2 Conﬂicts Detected

Conﬂicts are detected by combining the informa-

tion provided by the domain ontology (described by

the application designer) with the information from

WordNet database. Initially the domain-dependent

propagation occurs, considering relationships deﬁned

in the ontology. This step will create the following

norms:

Domain dependent propagation - Propagation

between entities:

(Contractor - Company 1)

hN1.1, F, B, Company 1, unaware(contract), , i

hN2.3, O, USA, Company 1, agree(contract), , i

hN3.4, F, U SA, Company 1, reproduce(tool), , i

hN4.5,F,USA,Company 1,commercialize(tool), , i

hN5.6,F,USA,Company 1,send(email), , i

hN6.7,O,USA,Company 1,purchase(software), , i

hN7.8,O,USA,Company 1,deliver (product), , i

hN8.9,F,USA,Company 1,access(server), , i

hN11.2,O,USA,Company 1,include (freight), , i

After propagation and domain-dependent proces-

sing, the original norms are joined to the propagated

norms and independent domain processing is perfor-

med. By using WorNet, the mechanism detects that

one of the contexts of the domain (California) is part

of other context of the domain (USA). Then, all norms

related to USA are always related to California. This

norm propagation which will result in the following

norms:

Domain-independent propagation - Propagation

between contexts:

(USA - California)

hN1.1,F,California,contractor,unaware(contract), , i

hN10.2,P,California,contracted,change(contract), , i

hN11.3,O,California,contractor,include(freight), , i

hN12.4,P,California,Company 1,unaware(clause), , i

hN13.5,P,California,Company 1,disagree(clause), , i

hN14.6,P,California,Company 1,copy(tool), , i

hN15.7,P,California,Company 1,login(computer), , i

hN17.8,F,California,contracted,edit(contract), , i

hN18.9,P,California,Company 1,send(spam), , i

hN19.10,P,California,contracted,exclude(product), , i

hN2.11,O,California,contractor,agree(contract), , i

hN20.12,F,California,Company 1,pay(software), , i

hN3.13,F,California,contractor,reproduce(tool), , i

hN4.14,F,California,contractor,commercialize(tool), , i

hN5.15,F,California,contractor,send(email), , i

hN6.16,O,California,contractor,purchase(software), , i

Table 2: Conﬂicts detected in the case study presented.

Conﬂict Conﬂict Pattern Norms

(N1,

N12)

-Time inter-

section

-Same Action

-Object Part-

Whole

(clause, contract)

- F x P

(hN1.1.20, F,California,

Company 1, unaware (contract),

, i,

hN12.4, P, California,

Company 1, unaware (clause), ,

(N2,

N13)

-Time inter-

section

-Action Anto-

nymy

(agree, disagree)

-Object Part-

Whole

(clause, contract)

-O x P

(hN2.3,O,USA,Company 1,

agree (contract), , i,

hN13,P,USA,Company 1,

disagree (clause), , i)

(hN2.3.22,O,California,

Company 1,agree(contract), , i,

hN13.5,P,California,

Company 1,disagree

(clause), , i)

(N3,

N14)

-Time inter-

section

-Action Syno-

nymy

(copy, reproduce)

-Same object

-F x P

(hN3.4,F,USA,Company 1,

reproduce(tool), , i,

hN14,P,USA,Company 1,

copy (tool), , i)

(hN3.4.23,F,California,

Company 1,reproduce (tool), , i,

hN14.6,P,California,

Company 1,copy(tool), , i)

(N5,

N18)

-Time inter-

section

-Same action

-Object Speciali-

zation

(spam, email)

-F x P

(hN5.6,F,USA,Company 1,

send(email), , i,

hN18,P,USA,Company 1,

send(spam), , i)

(hN5.6.25,F,California,

Company 1,send(email), , i,

hN18.9,P,California,

Company 1,send(spam), , i)

(N6,

N20)

-Time inter-

section

-Action Entail-

ment

(purchase, pay)

-Same object

-O x F

(hN6.7,O,USA,Company 1,

purchase(software), , i,

hN20,F,USA,Company 1,

pay(software), , i)

(hN6.7.26,O,California,

Company 1,purchase (soft-

ware), , i,

hN20.12,F,California,

Company 1,pay (software) , i)

(N9,

N19)

-Time inter-

section

-Action Reﬁne-

ment

(exclude, ma-

nage)

-Same object

-F x P

(hN9,F,USA,contracted,

manage(product), , i,

hN19,P,USA,contracted,

exclude(product), , i)

(N11,

N16)

-Time inter-

section

-Action Anto-

nymy

(exclude, include)

-Same object

-O x P

(hN11.2.21,O,California,

Company 1,include (freight), , i,

hN16,P,California,

Company

1,exclude(freight),

01/12/2016 00:00:00, 12/31/2016

00:00:00i)

hN7.17,O,California,contractor,deliver(product), , i

hN8.18,F,California,contractor,access(server), , i

hN9.19,F,California,contracted,manage(product), , i

hN1.1.20,F,California,Company 1,unaware(contract), , i

hN11.2.21,O,California,Company 1,include(freight), , i

hN2.3.22,O,California,Company 1,agree(contract), , i

hN3.4.23,F,California,Company 1,reproduce(tool), , i

hN4.5.24,F,California,Company 1,commercialize(tool), , i

hN5.6.25,F,California,Company 1,send(email), , i

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

hN6.7.26,O,California,Company 1,purchase(software), , i

hN7.8.27,O,California,Company 1,deliver(product), , i

hN8.9.28,F,California,Company 1,access(server), , i

Note that norms whose id is a triple numbering

(for instance, N5.6.25) are norms that were propa-

gated two times (domain-dependent propagation and

domain-independent propagation).

After the complete analysis, a total of 7 normative

conﬂicts were detected. It is important to emphasize

that conﬂicts can only occur between the norms that

are addressed to the same entities and contexts. Such

conﬂicts are listed in Table 2.

Note that the detection mechanism of our tool

combines two approaches, then, the conﬂict between

norms N9 and N19 was detected based on a domain

dependent relationship (relationship Reﬁnement bet-

ween actions). The remaining conﬂicts were detected

based on WordNet relationships involving actions and

objects (Synonymy, Antonymy, Entailment, Speciali-

zation, Part-Whole), and based on the relationship

Play speciﬁed on the domain ontology that determi-

nes that Company 1 plays the role of contractor.

6 CONCLUSIONS

The detection of conﬂicts among norms of a MAS is

a very important topic that has been studied in MAS.

When a system is governed by a large and/or com-

plex set of norms, a process of veriﬁcation/revision

of norms is needed to avoid the occurrence of norma-

tive conﬂicts. However, it is difﬁcult for a human to

detect all normative conﬂicts that may arise in a big

set of norms, besides being a task that demands a lot

of time and that is prone to errors. In this context,

our research aims to automate the conﬂict detection

process, providing a mechanism able to detect direct

and indirect normative conﬂicts. Our research provi-

des means to detect indirect normative conﬂicts that

do not depend on the domain of the application, that

is, it is able to detect conﬂicts even when the system

designer does not previously specify the relationships

among the elements of the norms, performing a se-

mantic mapping. In addition, we consider cases of

conﬂicts that may occur due to relationships that have

not yet been considered in any proposal of other aut-

hors, such as: synonyms and antonyms. However, we

know that a MAS can be designed to meet the needs of

a more speciﬁc universe, and therefore we cannot just

consider relationships that exist in the real world in

our analysis. For this reason, we have combined our

approach with the research presented in (Zahn, 2015)

and have developed a robust mechanism that can also

analyze relationships previously deﬁned in a domain

ontology. In short, our proposal to detect normative

conﬂicts was divided into three steps, as follows:

(i) the ﬁrst step consists in creating a mechanism

that maps the elements that compose the norms

to words and after that, search for relationships

between words. In this step, several cases of

normative conﬂicts were deﬁned that can be in-

ferred using WordNet as source of information.

This step is responsible for detecting domain-

independent conﬂicts;

(ii) the second step extends the approach presented

in (Zahn, 2015) so that it can be integrated with

the mechanism created in the ﬁrst step in order to

detect domain-dependent conﬂicts;

(iii) the third step consists of integrating the ﬁrst and

second steps and creating a tool for checking in-

direct conﬂicts involving relationships that de-

pend or not of the domain.

6.1 Limitations

Among the limitations of our research we can high-

light:

(i) the norms of the MAS must be described in an

OWL ontology following a speciﬁc format. We

consider this a limitation of our approach be-

cause if the system designer has a set of norms

described in natural language he will need to con-

vert this set of norms to the speciﬁc format deﬁ-

ned in the ontology;

(ii) domain-dependent relationships also need to be

described in an ontology and follow the format

detailed in (Zahn, 2015);

(iii) the developed tool is not able to distinguish

words that are homonymous, that is, words

whose spelling is the same, but have different

meanings when inserted in different contexts.

For example, the word doctor may refer to a per-

son graduated in medicine (medical context) or a

person who holds a doctorate degree (academic

context). This problem is known as Lexical Di-

sambiguation of Meaning (LDM) or Word Sense

Disambiguation (WSD) in Artiﬁcial Intelligence

(AI) (Ide and V

eronis, 1998). Although it is pos-

sible to analyze all the elements of a norm to try

to infer the meaning of a word in a given context,

this step is very costly and is not a guarantee of

certainty. In practice, such ambiguities are unli-

kely to occur during conﬂict detection because,

in general, norms are related to the same domain

or related ones. In addition, WordNet can store a

A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent Systems

word in different synsets that have a small diffe-

rence of meaning, which may make the process

of disambiguation unnecessary in many cases.

An example of this is the verb kill that belongs

to different synsets whose meanings are very si-

milar, as listed in Table 3.

Table 3: Meanings of the verb kill in WordNet.

(verb) kill

Gloss Example

causing death in-

tentionally

“This man killed several

people when he tried to

rob a bank”

be fatal “cigarettes kill”

deprive of life

“AIDS has killed thou-

sands in Africa”

causing death un-

intentionally

“She was killed in the

collision of three cars”

Despite these limitations, we consider that our re-

search provided a great contribution to the area of de-

tection of normative conﬂicts because it has resulted

in the creation of a tool that can help a software engi-

neer to design/include norms in a MAS in a consistent

way and mainly because it is possible to use such a

tool to detect normative conﬂicts in MAS that do not

have previously speciﬁed domain relationships. Thus,

the software engineer/system designer only needs to

specify in a domain ontology relationships that do not

occur in the real world.

6.2 Future Work

As suggestions for future work, we point out the fol-

lowing extensions for this research:

(i) to deﬁne cases of domain-independent conﬂicts

between norms that regulate states. This can be

implemented by using the majority of relations-

hips between deﬁned actions, but by referring to

another grammar class. To search for relations-

hips between actions we map actions to verbs

and to search for relationships between states we

could map states to adjectives;

(ii) to implement the lexical disambiguation process.

This process should consider the grammatical

class of the word and perform an analysis in-

volving all the elements that compose the norm

(context, entity, action, object). The disambi-

guation process can also investigate the seman-

tic similarity between two words, that is, can use

metrics to calculate numerical values that deter-

mine the proximity between a word and a con-

cept that contains a certain word. Other informa-

tion that may be useful for lexical disambiguation

are the words contained in the description (called

“gloss”) of each WordNet synset. In addition, if

the WordNet synset is associated with sentences

that exemplify the use of its words, such senten-

ces can also be used. However, since most ex-

isting LDM methods require high computational

complexity, many approaches adopt heuristics of

disambiguation, such as (Mihalcea, 2006):

a. the most common sense: words are disambi-

guated according to the most common sense of

the language. In this case, the algorithm should

select the ﬁrst synset that contains the word to

be disambiguated (because the WordNet sorts the

synsets according to their frequency of use);

b. a sense by discourse: after determining the me-

aning of a word p, all other occurrences of that

word will be attributed to the same meaning;

(iii) to receive as input of the algorithm norms in na-

tural language. In this case, it is necessary to de-

velop a mechanism for preprocessing a text ﬁle

that contain the norms in order to map them to the

norm deﬁnition presented in Section 3.1. Note

that when the norms are described in natural lan-

guage the modality of the norm (prohibition, per-

mission, obligation) may not be explicitly descri-

bed. To identify the modality of the norm, the

method can verify, for instance, if the norm has

some of the modal verbs, as follows:

a. verbs that denote obligations: must, need,

ought, have to, will;

b. verbs that denote permissions: can, may, could;

c. verbs that denote prohibitions (usually verbs de-

noting permissions and obligations accompanied

by a term denoting denial): cannot, should not,

must not, do not have.

REFERENCES

Aphale, M., Norman, T. J., and S¸ensoy, M., 2013. Goal-

directed policy conﬂict detection and prioritisation.

In Aldewereld, H., Sichman, J.S., (Eds), Coordina-

tion, organisations, institutions and norms in agent sy-

stems VIII, volume 7756 of Lecture notes in computer

science (pp. 87104). Springer.

S¸ensoy, M., Norman, T. J., Vasconcelos, W. W., and Sycara,

K., 2012. OWL-POLAR: A framework for semantic

policy representation and reasoning. Web Semantics:

Science, Services and Agents on the World Wide Web,

1213, 148160.

Fenech, S., Pace, G. J., and Schneider, G., 2008. Detection

of conﬂicts in electronic contracts. NWPT 2008, 34.

Gaertner, D., Garcia-Camino, A., Noriega, P., Rodriguez-

Aguilar, J. A., and Vasconcelos, W. W., 2007. Distri-

buted norm management in regulated multiagent sy-

stems. In Proceedings of the 6th international joint

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

conference on autonomous agents and multiagent sy-

stems, AAMAS 07 (pp. 90:190:8). New York, NY:

ACM.

Giannikis, G. K. and Daskalopulu, A., 2011. Normative

conﬂicts in electronic contracts. Electronic Commerce

Research and Applications, 10(2), 247267.

Horridge M. and Bechhofer S., 2011. The OWL api: A java

api for owl ontologies. Semantic Web, v. 2, n. 1, p.

11-21.

Ide, N. and V

eronis, J., 1998. Introduction to the special

issue on word sense disambiguation: the state of the

art. Computational linguistics, v. 24, n. 1, p. 2-40.

Kollingbaum, M. J., Norman, T. J., Preece, A., and

Sleeman, D., 2007. Norm conﬂicts and inconsisten-

cies in virtual organisations. In P. Noriega, J. V

azquez-

Salceda, G. Boella, O. Boissier, V. Dignum, N. For-

nara, & E. Matson (Eds.), Coordination, organizati-

ons, institutions, and norms in agent systems II, vo-

lume 4386 of Lecture notes in computer science (pp.

245258). Berlin: Springer.

Mihalcea, R., 2006. Knowledge-based methods for WSD.

Word Sense Disambiguation: Algorithms and Appli-

cations, pp. 107-131.

Miller, G. A., 1995. WordNet: a lexical database for Eng-

lish, Communications of the ACM, 38(11), 3941.

Neto, B. F. S., Silva, V. T., and Lucena, C. J. P., 2012. An

architectural model for autonomous normative agents.

In L. Barros, M. Finger, A. Pozo, G. Gimen

enez-

Lugo, & M. Castilho (Eds.), Advances in artiﬁcial

intelligenceSBIA 2012, Lecture notes in computer

science (pp. 152161). Berlin: Springer.

Neto, B. F. S., Silva, V. T., and Lucena, C. J. P., 2013. De-

veloping goal-oriented normative agents: The NBDI

architecture. In J. Filipe, & A. Fred (Eds.), Agents and

artiﬁcial intelligence, volume 271 of Communications

in computer and information science (pp. 176191).

Berlin: Springer.

Santos, J. S., and Silva, V. T., 2016. Identifying Domain-

Independent Normative Indirect Conﬂicts. In: 2016

IEEE 28th International Conference on Tools with Ar-

tiﬁcial Intelligence (ICTAI), 2016, San Jose. p. 536.

Santos, J. S., Zahn, J. O., Silvestre, E. A., Silva, V. T.,

and Vasconcelos, W. W., 2017. Detection and reso-

lution of normative conﬂicts in multi-agent systems:

a literature survey. Autonomous Agents and Multi-

Agent Systems, pp. 1-47. DOI: [ONLINE] DOI:

10.1007/S10458-017-9362-Z.

Vasconcelos, W. W., Kollingbaum, M. J., and Norman,T. J.,

2009. Normative conﬂict resolution in multiagent sys-

tems. Autonomous Agents and Multi-Agent Systems,

19(2), 124152.

Vasconcelos, W. W., Garc

ıa-Camino, A., Gaertner, D.,

Rodr

ıguez-Aguilar, J. A., and Noriega, P., 2012. Dis-

tributed norm management for multi-agent systems.

Expert Systems with Applications, 39(5), 5990 5999.

Walenz, B. and Didion, J., 2011. “JWNL: Java WordNet li-

brary.” 2008-05-14]. http://jwordnet. sourceforge, net.

Zahn, J. O., 2015. Um Mecanismo de Veriﬁcac¸

ao de Con-

ﬂitos Normativos Indiretos. Masters thesis, Instituto

de Computac¸

ao - Universidade Federal Fluminense

(IC/UFF), Niter

oi, Brasil.

A Novel Tool for Detecting Indirect Normative Conﬂicts in Multi-agent Systems