SOURCE SENSITIVE ARGUMENTATION SYSTEM

Chee Fon Chang, Peter Harvey and Aditya Ghose

Decision Systems Lab

University of Wollongong, NSW 2522, Australia

Keywords:

Multi-agent Argumentation, Multi-agent Negotiation.

Abstract:

There exist many approaches to agent-based conﬂict resolution. Of particular interest are approaches which

adopt argumentation as their underlying conﬂict resolution machinery. In most argumentation systems, the

argument source plays a minimal role. We feel that ignoring this important attribute of human argumentation

reduces the capabilities of current argumentation systems. This paper focuses on the importance of sources in

argumentation and extends this to the notion of credibility of agents.

1 INTRODUCTION

There have been many recent developments on logical

systems for defeasible argumentation and the use of

argumentation in multi-agent interaction such as ne-

gotiation, group decision making and dispute media-

tion. Systems such as (Prakken, 1993; Dung, 1995;

Bondarenko et al., 1997; Vreeswijk, 1997; Dung and

Son, 2001), provides formalisation to defeasible or

non-monotonic reasoning. These systems focus on

representation and general interaction that exist be-

tween arguments. (Jennings et al., 1998; Sierra et al.,

1997; Kraus et al., 1998; Parsons et al., 1998), studies

argumentation as a conﬂict resolution machinery of

negotiation, where arguments for an offer should per-

suade the other party to accept the offer. These recent

applications of argumentation in multi-agent systems

have drawn great interest. If one was to use defeasi-

ble argumentation as the conﬂict resolution machin-

ery, modiﬁcations are required.

We view argumentation as begin (simultaneously)

a process for information exchange, a process for

conﬂict and an approach to knowledge representa-

tion/reason. In (Kraus et al., 1998), the authors stated

that argumentation is a process that has the struc-

ture of a logical proof however does not hold the

strength of one. Consistent with this view, we believe

that multi-agent argumentation focuses not on logical

“truth” but on convincing/persuading other agents of

a particular view or position. As such, how an indi-

vidual is perceived in their community will effect the

acceptance of their arguments.

In the next section, we will motivate this work with

a simple example. In section 2, a formalisation of

source sensitive argumentation system is provided.

Section 3 provides the procedures and worked exam-

ples. Section 4 describes the results and behavioural

impacts.

1.1 Motivation

We will ﬁrst motivate our work using an adaptation of

an example from (Verheij, 1994) and then extend the

example to capture real-life situations. Let’s assume

that Bill is a juvenile and has committed a crime. We

pose a question: “Should Bill be punished?”. Assum-

ing that you are given the statements below:

Bill has robbed someone, so he should be jailed.(α)

Bill is a juvenile, therefore he should not go to jail.(β)

In the above statement, if we assume that the sec-

ond statement is stronger, then we can say that β

defeats α. However, the above form of argumenta-

tion is insufﬁcient in capturing real-life argumenta-

tion. Existing argumentation systems (Jennings et al.,

1998; Parsons et al., 1998; Sierra et al., 1997; Kraus

et al., 1998; Dung, 1995; Verheij, 1994; Pollock,

1991; Vreeswijk, 1997) separate the source of the ar-

gument from the argument when evaluating the de-

feat. We believed that the validity and strength of an

argument cannot be capture by the argument alone. It

is common in reality for one to evaluate the strength

Fon Chang C., Harvey P. and Ghose A. (2006).

SOURCE SENSITIVE ARGUMENTATION SYSTEM.

In Proceedings of the Eighth International Conference on Enterprise Information Systems - AIDSS, pages 39-46

DOI: 10.5220/0002464800390046

 SciTePress

and validity of arguments with respect to the provider

of the argument. Furthermore, we claim that “under-

cut”(Pollock, 1970) (a form of attack) cannot be per-

formed without this meta information.

Consider the following modiﬁcation of the above

example:

Tom: Bill has robbed someone, so he should be

jailed.(α)

Dick: Bill is a juvenile, therefore he should not go to

jail.(β)

Previously, we assumed that the second statement

is stronger and hence we said that β defeats α.If

we knew that Tom is a known liar it would further

strengthen the acceptability of this defeat. However,

if you were told or knew that Tom is a police ofﬁcer

we would be less willing to accept the defeat.We

would most likely rule in favour of Tom.

We will now extend this example further. Let us as-

sume that Tom is a police ofﬁcer, however over a pe-

riod of time allegations of wrongful arrest, brutality

and violation of arrest procedures against Tom were

made by an independent source. These allegations

would diminish Tom’s credibility. Now, if we were

asked to evaluate subsequent arguments between Tom

and Dick or to re-evaluate the defeat between α and

β the original defeat may no longer hold. From the

above examples, it is clear that the defeat relation can-

not simply be between arguments.

Proposals by (Carbogim, 2000) capture the affect

of the underlying knowledge base by arguments, how-

ever the proposal does not deal with changing defeat

relation. In this paper, we will show that by simply by

tagging arguments with sources, the resulting defeat

relation is more ﬂuent and changes the dynamics of

argumentation systems.

2 FORMAL DEFINITION

As stated in (Prakken and Vreeswijk, 2002), any de-

feasible argumentation system consists of several im-

portant components: a language, a deﬁnition of argu-

ments, a deﬁnition of inconsistency/conﬂict/attack, a

defeat relation and ﬁnally, the overall status of the ar-

gument. We will now provide the components of our

system closely matching these key components.

For simplicity, we will take any ﬁnitely generated

propositional language L with the usual punctuation

signs, and logical connectives ¬ (not) and ⊃ (im-

plies). For any set of wffs S ⊆Land any α ∈L,

S  α means α is provable from premises S. For any

set of wffs S ⊆L,Cn

(S) = { α | S  α }.

Deﬁnition 1 (Argument) An argument α is a triple

F, A, C where F, A and C denote the sets of facts,

assumptions and conclusion respectively.

Deﬁnition 2 (Well-founded Argument) An argument

α is a well-founded argument iff it satisﬁes the fol-

lowing conditions:

• F , A, C ⊆L

• F ∩ A = ∅

• F ∪ A  C

• Cn(F ∪ A ∪ C) ⊥

We will write F

, A

and C

to respectively de-

note the facts, assumptions and conclusions associ-

ated with an argument α. One can view a well-

founded argument as a premise-conclusion pair where

the premise is F, A and the conclusion is C. By al-

lowing assumptions, this generalise the arguments to

allow for representation of weak facts. Note that since

we are interested in rational agents, we have elim-

inated self-defeating (Prakken and Vreeswijk, 2002;

Pollock, 1991) arguments. If one was interested in

creating irrational agents or paradoxes, one can make

the appropriate changes by removing the last condi-

tion.

Deﬁnition 3 (Conﬂict) A pair of well-founded ar-

guments β and γ are said to be in conﬂict iff

Cn(C

∪ C

) ⊥

Deﬁnition 4 (Attack) For a pair of well-founded ar-

guments β and γ we say that β attacks γ iff:

1. β and γ are in conﬂict

2. Cn(F

∪ F

∪ A

) ⊥

There exists great variation of the deﬁnition for

conﬂict/attack in the defeasible argumentation liter-

ature, and are often deﬁned a way that makes them

interchangeable. In our system, these notions are in-

stead assigned speciﬁc meanings.

Conﬂict represents an inconsistency between the

conclusion and is symmetrical. Attack represents an

inconsistency on the premise. It is our view that if

the conclusions are non-conﬂicting then there are no

issues to be resolved. Generally, if we are unaware

of a difference in opinion, we do not argue about it.

Later we will combine these two notions to deﬁne re-

buttal, assumption attack and undercut. Note that our

deﬁnitions will differ slightly from those described in

(Prakken and Vreeswijk, 2002).

Deﬁnition 5 (Tagged Arguments) Given a set of

unique identiﬁers I, we deﬁne A as a set of tagged

arguments of the form S, A where

• S ∈Irepresents the tagged arguments’ source.

• A is a set of well-founded arguments.

We will write S

and A

to respectively denote the

source and well-founded argument associated with a

tagged argument φ.

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

Deﬁnition 6 (Credibility Function) Given a set of

unique identiﬁers I, we say C is a credibility function

if it maps all values of I into a some totally ordered

set.

The notion of credibility provides the agent with a

measure of strength belief per argument source. This

notion also provides a simplistic measure of trust. For

simplicity, we have deﬁne it as a function that maps a

set of unique identiﬁers into a total ordered set. How-

ever, one could deﬁne an arbitrary complex measure

taking into account of the context in which the argu-

ment is situated. For example, if the agent is a stock-

broker, then any arguments related to the share market

from this agent will be more credible than any agent

that is not a stock-broker. One could also deﬁne a

mapping to an arbiter set as long as there exist oper-

ators that are transitive and asymmetric deﬁned over

the set.

Deﬁnition 7 (Defeat) Given a set of tagged argu-

ments A and the credibility function C, a relation

D ⊆A×Ais said to be a defeat relation on A.

We will write φDψ iff at least one of the following if

true:

• A

attacks A

and A

does not attack A

• A

attacks A

and C(S

) >C(S

)

• A

and A

are in conﬂict, neither A

attacks A

nor A

attacks A

and C(S

) >C(S

)

The tagging of arguments allows us to uniquely

identify the argument source, and so we make use of

this in our deﬁnition of defeat. Also, note that our no-

tion of defeat is not deﬁned as a global relation but

as a per-agent defeat relation and is determined by a

credibility function C (see later deﬁnition 8).

Our deﬁnition of defeat also encapsulates vari-

ous types of defeat. For example, (Prakken and

Vreeswijk, 2002) states that assumption attack oc-

curs when one argument proves what was assumed

unprovable by the ﬁrst (in other words, when a con-

clusion of one argument attacks the assumption of

another). We will say that assumption attack occurs

when facts of one argument attacks the assumption of

another argument. This is captured by attack(φ, ψ)∧

¬attack(ψ,φ).

Similarly, (Prakken and Vreeswijk, 2002) states

that rebuttal occurs when the conclusion of one ar-

gument attacks the premise of another. We deviate

slightly and say that rebuttal occurs when facts of one

argument attack facts of another argument. This is

captured by attack(φ, ψ) ∧ C(S

) >C(S

Finally, (Prakken and Vreeswijk, 2002) states that

undercut occurs when an argument attacks an infer-

ence rule used in another argument. In our sys-

tem undercut occurs when the conclusion of the

arguments contradict but neither the facts nor as-

sumptions of both arguments contradict, as this

implies that the contradiction occurs during infer-

ence. This notion is capture by conf lict(φ, ψ) ∧

¬ (attack(φ, ψ) ∨ attack(ψ, φ)) ∧ C(S

) >C(S

)

Deﬁnition 8 (Agent) Given a set of unique identiﬁers

I and a set of tagged arguments A, an agent is repre-

sented as a 4-tuple of the form I, A, C where

• I ∈I.

• A ⊆As.t. ∀φ : φ ∈A,ifS

= I then φ ∈ A

• C is a credibility function. This represents the cred-

ibility (higher values are better) of other agents in

the system as evaluated by the agent.

Note that the credibility function is subject to revi-

sion by the agent during the execution of the system

as each agent adjusts it’s view of fellow agents. Note

also that the set of tagged arguments A is subjected

to change as individual agent discover new arguments

during the argumentation process. We do not require

the agent to know all arguments nor do we require

that the set of arguments be conﬂict-free. The notion

of conﬂict-free set is simply no two arguments in the

set defeat each other.

The uniqueness of this system is that there ex-

ists no global consensus on the credibility value on

each agent. This measure is recorded from individual

agent’s perspective and is stored by individual agents.

Given that there is no requirement for global consen-

sus on individual agent’s credibility, consensus on the

amount of adjustment on the credibility value is not

required either. However, for agents to be produc-

tive, we believed that there should be a consensus on

when and the kind of adjustment that should be per-

formed. A simple rule would be that if an observation

is made of an agent winning an argument, then that

agent’s credibility should be adjusted upwards and the

converse holds. One could also extend this to cap-

ture situation importance. For example, if an agent

is observed to have won an important argument, then

it’s credibility is revised upwards by a greater value

to that of an argument with less important. Similar

to human debates, this provides a notion of a “career

making win”. This also provides incentive for agents

to win the argument. We have provided directions in

which one could formula a true-to-life function, we

leave the function details to designers.

For convenience we will write D

to denote the de-

feat relation (from deﬁnition 7) as determined by the

A and C held by agent ρ.

Deﬁnition 9 (Stable) For an agent ρ, we say a set

S ⊆ A is a stable set of arguments for that agent iff

it is a maximal (w.r.t. set inclusion) set of arguments

such that:

•∀ψ ∈ A

− S, ∃φ ∈ S that conﬂicts with ψ.

•∀φ ∈ A

, ψ ∈ S where ψD

φ.

SOURCE SENSITIVE ARGUMENTATION SYSTEM

A stable set provides a particular view which the

agent has adopted. To select a particular view is to

side with some set of agents in the system hence pro-

viding support to that group of agents, in essence

forming parties or coalitions. We will write S

to de-

note a stable set of arguments adopted by an agent ρ.

Deﬁnition 10 (Source Sensitive Argumentation Sys-

tem) A source sensitive argumentation system is de-

ﬁned as:

SAS = Agt, A

where

• Agt is a set of agents.

•Ais a set of tagged arguments.

Deﬁnition 11 (Consensus) Given a source sensitive

argumentation system SAS = Agt, A, we say a set

S⊆Ais a consensus iff for all agents ρ ∈ Agt, there

exist a state S

such that no argument in S

argument in S.

A consensus is a set of arguments that cannot be

defeated by any agent.

Deﬁnition 12 (G-consensus) Given a source sensi-

tive argumentation system SAS = Agt, A and

G ⊆ Agt, we say a set S⊆Ais a G-consensus iff

for all agents ρ ∈ G, there exist a state S

such that

no argument in S

an argument in S.

We understand that it is generally very hard to reach

a consensus. G-consensus can be viewed as group

consensus where a predeﬁned set of agents are con-

sulted when determine consensus. So if G is the com-

plete set of the agents in the system then G-consensus

is equal to consensus.

Deﬁnition 13 (N-acceptable) Given a source sensi-

tive argumentation system SAS =  Agt, A, we say

a set S

⊆Ais a n-acceptable iff there exists some

G ⊆ Agt such that |G| = n and S

is a G-consensus.

With n-acceptable, we simply denote the number of

agents that we will seek consensus and not any spec-

ify a group.

Deﬁnition 14 (Majority) Given a source sensitive ar-

gumentation system SAS = Agt, A, we say a

non-empty set S

⊆Ais a majority if it is

|Agt|

acceptable.

Majority is simply a set of arguments that more

than half the agent population are happy with.

3 PROCEDURE

Informally, an agent’s belief consists of a set of ar-

guments, and a credibility ordering derived from the

credibility function. Each agent will revise their

credibility function as the argumentation process pro-

gresses, and so we assume that each agent has a cred-

ibility revision operator. We leave the details of the

revision operator to the individual agent’s designer,

though we note that it is possible for an agent to as-

sign other agents higher credibility than itself.

For the purpose of our discussion, we propose the

use of the following basic revision rules:

• If an agent is observed to have lost an argument, its

credibility is reduced.

• If an agent is observed to have won an argument,

its credibility is increased.

• We revise only when a new observation is made,

and so do not repeatedly reduce/increase the credi-

bility of an agent.

We note that the above general rules still permit

agents to participate in more than one argumentation

session.

We will also assume that there exists a communi-

cation protocol used by agents to transmit and receive

arguments. For simplicity, we will assume that the ar-

guments transmitted are transparent to all agents. As

a simplistic protocol to the argument exchange, we

suggest that agents transmit their argument in half-

duplex mode and synchronously, hence taking turns

in presenting their argument or counter-argument. At

the end of each round of exchange, individual agents

will determine the defeat of relationship between ar-

guments. The agents will then present their outcome

to so that the consensus and majority can be com-

puted. Following this simple procedure we will now

provide an example of multi-agent dialogue that is in

line with our argumentation system.

We introduce three agents to frame our example;

Tom, Dick and Harry. Tom (a police ofﬁcer) holds

the argument α, which is ‘Bill has robbed someone,

so he should be jailed’. Dick will hold the argument

β, which is ‘Bill is a juvenile, therefore he should not

go to jail’. Harry holds no argument yet, but will par-

ticipate in the argumentation process. Therefore ini-

tially:

Beliefs At Start

Tom C(Tom) >C(Dick) >C(Harry)

{α} is the outcome

Dick C(Dick) >C(Tom) >C(Harry)

{β} is the outcome

Harry C(Harry) >C(Tom) >C(Dick)

{} is the outcome

At this initial point there is no consensus between

the agents, and so S = ∅. Similarly, there is no ma-

jority, and so S

= ∅. Note that Harry sees Tom (the

police ofﬁcer) as more credible than Dick.

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

We assume now that Tom and Dick communicate

their arguments to the other agents, and so all agents

know the arguments Tom,α and Dick,β. This

gives the following situation:

Beliefs After Communicating Arguments

Tom C(Tom) >C(Dick) >C(Harry)

Tom,α defeats Dick,β

{α} is the outcome

Dick C(Dick) >C(Harry) >C(Tom)

Dick,β defeats Tom,α

{β} is the outcome

Harry C(Harry) >C(Tom) >C(Dick)

Tom,α defeats Dick,β

{α} is the outcome

There remains no consensus between the agents,

and so S = ∅. However, there is now a majority,

and so S

= {Tom,α}.

We now consider what is to happen if the reputa-

tion of Tom was tarnished. Formally, this would mean

that Harry changes his assessment of the credibility of

each agent so that C(Harry) >C(Dick) >C(Tom).

Note that this could happen through some other argu-

mentation process involving Dick and Harry.

Beliefs After Decreasing Tom’s Credibility

Tom C(Tom) >C(Dick) >C(Harry)

Tom,α defeats Dick,β

{α} is the outcome

Dick C(Dick) >C(Harry) >C(Tom)

Dick,β defeats Tom,α

{β} is the outcome

Harry C(Harry) >C(Dick) >C(Tom)

Dick,β defeats Tom,α

{β} is the outcome

Finally, we consider the situation where Tom him-

self is ‘ashamed’ and lowers his own credibility be-

low that of Dick’s. As shown below, a consensus is

established, with all agents agreeing on β. From this

example we can see that consensus will only be es-

tablished if there was no disagreement, or one agent

‘gives ground’ and lowers their own credibility. We

will explore this more completely below.

Beliefs After Tom ‘Gives Ground’

Tom C(Dick) >C(Tom) >C(Harry)

Dick,β defeats Tom,α

{β} is the outcome

Dick C(Dick) >C(Harry) >C(Tom)

Dick,β defeats Tom,α

{β} is the outcome

Harry C(Harry) >C(Dick) >C(Tom)

Dick,β defeats Tom,α

{β} is the outcome

3.1 Cycles

There exists two forms of cycles. Firstly, cycles on

the defeat relation between arguments. This is caused

by agents ranking themselves as deﬁnitive source of

credibility. Below is an example of a cycle.

Cycle in Defeat Relation

Tom Tom,α defeats Dick,β

C(Tom) >C(Dick))

{α} is the outcome

Dick Dick,β defeats Tom,α

C(Dick) >C(Tom)

{β} is the outcome

This form of cycle results in no consensus and, as

in this case, may result in no majority. This form of

cycle is not a serious issue as it indicates to the oper-

ators that no agreement can be reached, but does not

hinder the termination of the program. Note also that

we do expect such cycles to occur, as they result from

underlying cycles in the attacks relation.

The second form of cycle occurs during the com-

putation of consensus and majority. This is caused by

agents inability to commit to a particular credibility

ordering. This is directly attributed to a bad revision

operator. We will term this type of cycle oscillation.

For the purpose of this discussion, we will assume

that the revision operators deﬁned by designers are

well behaved. With a well-behaved revision opera-

tor, cycles of random raising and lowering credibility

never occur. Furthermore, after winning an argument

an agent is more likely to win future arguments. This

self-supporting nature further ensures that cycles of

raising and lowering credibility never occur. This en-

sures that a ﬁx-point is achieved, and hence a termi-

nation condition for the program exists.

SOURCE SENSITIVE ARGUMENTATION SYSTEM

4 RESULTS AND BEHAVIOURAL

IMPACTS

This section will outline the contributions made by

this system. Agent behaviours resulting from this sys-

tem falls into the categories of: agent reasoning, agent

autonomy and group behaviours.

4.1 Agent Reasoning

In (Jennings et al., 1998; Sierra et al., 1997; Parsons

et al., 1998), the authors proposed the use of argu-

mentation to assist in agent negotiation. Information

that are naturally embedded in arguments are use to

determine or eliminate infeasible proposals. With the

additional notion of credibility, our system further as-

sist in this process. With an embedded defeat relation,

agents are in the position to evaluate their own argu-

ments before communicating it to other agents. This

creates the notion of a minimal winning argument. A

minimal winning argument increases its’ strength by

minimising the possible attack on it. For example, by

reducing the number of assumptions in an argument,

minimises the possible assumption attack. The notion

of a minimal argument allows for the arguments to

be ranked according to strength and provides an or-

der in which arguments are to be presented to fellow

agents. The introduction of credibility also introduces

a notion of self-preservation. If an agent wishes to be

in a position to win future arguments, then it must

keep its’ credibility value high. The self-preservation

notion forces the agent into evaluate arguments more

carefully so as not to randomly propose weak argu-

ments. Essentially agents be will force into a position

in which they only “ﬁght in battles that they believe

they can win”. This improves the efﬁciency of the

argumentation process.

Additionally, agents are now equipped with addi-

tional rationale. The use of credibility provides an-

other means to support their decisions.

4.2 Agent Autonomy

The lack of a global defeat relation and consensus on

the credibility provides this system the ability to dis-

tribute and scale. This feature do away with the need

for a central repository/location where conﬂicts are

resolved. This system also allows for arguments to

propagate hence conﬂicts can be isolated and resolved

independently. This approach also removes the notion

of a static top argument.

4.3 Agent Group Behaviour

Group decision making and coalition formation have

received huge amount of interest recently. Our sys-

tem provides agents with the ability to support fellow

agents. This is archived through the consensus and

majority operation. If an agent’s argument is deem

strong, it would persuade fellow agents of its’ view.

This exercise causes parties to form and hence coali-

tion to form leading to dynamic coalition formation.

5 RELATED WORKS

This section will outline the related works. The re-

lated works falls into three categories: defeasible

argumentation, argumentation-based negotiation and

distributed constraints satisfaction.

5.1 Defeasible Argumentation

Although the works such as (Dung, 1995; Prakken,

1993; Bondarenko et al., 1997; Dung and Son, 2001;

Dung, 1995; Prakken, 1993; Vreeswijk, 1997) fo-

cuses on representation and logical reasoning, we

have shown that by introducing the notion of sources

and credibility as well as a corresponding notion of

defeat, we are able to provide a mapping for a theo-

retical to a practical system. These systems provide

a strong basis for our proposal. We have modelled

our system around these collection of defeasible argu-

mentation systems. We feel that our work will com-

plement those advances already achieved.

Additionally, recent work (Prakken, 2001; Car-

bogim, 2000; Brewka, 2001) in defeasible argumen-

tation have focused on the notion of dynamic argu-

mentation. (Prakken, 2001) focus on the fairness

and soundness of dynamic argumentation protocol.

(Carbogim, 2000) focused on addressing issues of

associated to change in the underlying knowledge

base caused by arguments. (Brewka, 2001) deals

with meta-level argumentation, providing undercut

via a notion of preference on which defeat rule holds.

Components of these work have similarities to ours,

however the notion of sources and credibility forcing

a change of preference on the defeat rules are not in-

vestigated. In fact these works do not have explicit

notion of sources and credibility.

5.2 Argumentation-Based

Negotiation

The exchange of arguments and counterarguments

has also been studied in the context of multi-agent in-

teraction. In (Kraus et al., 1998) and (Parsons et al.,

1998) have studied argumentation as a component of

negotiation protocols, where arguments for an offer

should persuade the other party to accept the offer.

Generally, these works focus on the negotiation pro-

tocol and agent behaviours, leaving the representation

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

and internal reasoning to the designer. In (Kraus et al.,

1998), a range of agents and behaviours were pre-

scribed. We feel that our work further enhance the

internal reasoning of these agents. Although some

notion of sources and credibility may have existed in

(Jennings et al., 1998; Sierra et al., 1997), these no-

tions were not explicitly utilised when evaluating ar-

guments.

5.3 Distributed Constraint

Satisfaction

Recent work in distributed constraint satisfaction al-

gorithms (Harvey et al., 2005b; Harvey et al., 2005a;

Harvey et al., 2006) is built upon the theoretical un-

derpinnings described in this paper. Support-Based

Distributed Search (SBDS) is a distributed constraint

satisfaction algorithm in which agents communicate

via arguments, maintaining a simple notion of credi-

bility between agents.

The argument structure of SBDS is domain spe-

ciﬁc, permitting two categories of arguments. As it

deviates from the argumentation structure representa-

tion in this paper (with distinction between facts and

assumptions). Below is a loose description of the ar-

gument structure below in SBDS.

Deﬁnition 15 (SBDS Argument) An SBDS-

argument is a pair Prem,Con, and belongs to one

of the following two categories:

1. isgoods (variable-value assignment proposals)

• Prem - an ordered sequence of variable-value as-

signments

• Con - the variable-value assignment for the agent

stating the argument

2. nogoods (variable-value assignment rejections)

• Prem - a set of variable-value assignments which

are not permitted

• Con - exact copy of the premise

As the argument structure of SBDS differ from

the monotonic logic example given in this paper, a

domain-speciﬁc conﬂict and attacks relation was de-

ﬁned. The spirit of these relation remains the same

as those provided in our work. This further strength-

ens our claim that our proposal complements and pro-

vides assistance to solving problems in a whole range

of different domains.

6 CONCLUSION

In most argumentation systems, the source of the ar-

gument plays a minimal role. In this paper, we have

introduced the notion of sources and credibility. We

have shown that by simply tagging arguments with

sources, the resulting defeat relation is more ﬂuent

and the dynamics of argumentation systems changes.

We have shown that ignoring this important attribute

of human argumentation reduces the capabilities of

current argumentation systems. This paper focuses

on the importance of information source in argumen-

tation and extending this to the notion of credibility.

The notion of credibility plays an important role in

the agent decision making process during argumenta-

tion. We have also shown that the system is capable of

emulating feature of human argumentation that was

not captured in existing systems. Finally, we have

shown that this system is not purely theoretical but

can also be applied into the practical domain such as

argumentation-based negotiation and distributed con-

straint satisfaction.

6.1 Future Works

Throughout this paper, we have indicated room for

improvements. We point to improvements that can be

made on mapping sources to credibility by augment-

ing the function to consider context. We also like to

point out that the current mapping may not be satis-

factory to some audience. Suggestion of generalising

the mapping function via the use of semi-ring struc-

ture, combining two orthogonal metric (one measur-

ing strength of the argument, the other measuring the

credibility) are currently under investigation. By us-

ing this approach, we are in the position to provide

a notion of graded defeat. This modiﬁcation would

provide a method to infer the global credibility and

defeat relation from a given set of agents.

REFERENCES

Bondarenko, A., Dung, P. M., Kowalski, R. A., and Toni,

F. (1997). An abstract, argumentation-theoretic ap-

proach to default reasoning. Artiﬁcial Intelligence,

93:63–101.

Brewka, G. (2001). Dynamic argument systems: A formal

model of argumentation processes based on situation

calculus. Journal of Logic and Computation, 11.

Carbogim, D. V. (2000). Dynamics in Formal Argumenta-

tion. PhD thesis, University of Edinburgh. College of

Science and Engineering. School of Informatics.

Dung, P. M. (1995). An argumentation-theoretic founda-

tions for logic programming. Journal of Logic Pro-

gramming, 22(2):151–171.

Dung, P. M. and Son, T. C. (2001). An argument-based

approach to reasoning with speciﬁcity. Artiﬁcial In-

telligence, 133(1-2):35–85.

Harvey, P., Chang, C. F., and Ghose, A. (2005a). Practi-

cal application of support based distributed search. In

Proceedings of the 17th IEEE International Confer-

ence on Tools with AI.

SOURCE SENSITIVE ARGUMENTATION SYSTEM

Harvey, P., Chang, C. F., and Ghose, A. (2005b).

Support-based distributed search. In Meisels,

A., editor, Proceedings of the 6th International

Workshop on Distributed Constraint Reasoning.

http://www.dsl.uow.edu.au/people/harvey/dcr05.pdf.

Harvey, P., Chang, C. F., and Ghose, A. (2006). Support-

based distributed search. In Proceedings of the 5th

International Joint Conference on Autonomous Agents

& Multi Agent Systems.

Jennings, N. R., Parsons, S., Noriega, P., and Sierra, C.

(1998). On argumentation-based negotiation. In

Proceedings of the International Workshop on Multi-

Agent Systems, Boston, USA.

Kraus, S., Sycara, K. P., and Evenchik, A. (1998). Reaching

agreements through argumentation: A logical model

and implementation. Artiﬁcial Intelligence, 104(1-

2):1–69.

Parsons, S., Sierra, C., and Jennings, N. R. (1998). Agents

that reason and negotiate by arguing. Journal of Logic

and Computation, 8(3):261–292.

Pollock, J. L. (1970). The structure of epistemic justiﬁca-

tion. American Philosophical Quarterly, monograph

series 4:62–78.

Pollock, J. L. (1991). Self-defeating arguments. Minds and

Machines, 1(4):367–392.

Prakken, H. (1993). An argumentation framework in

default logic. Annals of Mathematics and Ar-

tiﬁcial Intelligence, 9(1-2):93–132. Bibsource =

DBLP,http://dblp.uni-trier.de.

Prakken, H. (2001). Relating protocols for dynamic dispute

with logics for defeasible argumentation. Synthese,

127(1):187–219.

Prakken, H. and Vreeswijk, G. (2002). Logics for Defeasi-

ble Argumentation, volume 4 of Handbook of Philo-

sophical Logic, pages 218–319. Kluwer Academic

Publishers, second edition edition.

Sierra, C., Jennings, N. R., Noriega, P., and Parsons, S.

(1997). A framework for argumentation-based negoti-

ation. In Singh, M. P., Rao, A. S., and Wooldridge, M.,

editors, Proceedings of the 4th International Work-

shop on Agent Theories, Architectures, and Lan-

guages, pages 177–192.

Verheij, B. H. (1994). Reason based logic and legal knowl-

edge representation. In Carr, I. and Narayanan, A.,

editors, Proceedings of the 4th National Conference

on Law, Computers and Artiﬁcial Intelligence, pages

154–165. University of Exeter.

Vreeswijk, G. (1997). Abstract argumentation systems. Ar-

tiﬁcial Intelligence, 90(1-2):225–279.

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS