Agents that learn how to generate arguments from other agents
1
Agents that learn how to generate
arguments from other agents
Ariel MONTESERIN and Analía AMANDI
ISISTAN Research Institute, CONICET - UNCPBA, Campus Uni-
versitario, Tandil, Bs. As. Argentina
ariel.monteserin, analia.amandi@isistan.unicen.edu.ar
Received 11 October 2012 - Revised 29 April 2013
Abstract
Learning how to argue is a key ability for a negotiator agent.
In this paper, we propose an approach that allows agents to learn how
to build arguments by observing how other agents argue in a negotiation
context. Particularly, our approach enables the agent to infer the rules
for argument generation that other agents apply to build their arguments.
To carry out this goal, the agent stores the arguments uttered by other
agents and the facts of the negotiation context where each argument is
uttered. Then, an algorithm for fuzzy generalized association rules is ap-
plied to discover the desired rules. This kind of algorithm allows us (a) to
obtain general rules that can be applied to dierent negotiation contexts;
and (b) to deal with the uncertainty about the knowledge of what facts
of the context are taken into account by the agents. The experimental
results showed that it is possible to infer argument generation rules from
a reduced number of observed arguments.
Keywords
Intelligent Agents,
Argumentation-based Negotiation,
Fuzzy Generalized Association Rules,
Learning.
1 Introduction
2
Ariel MONTESERIN and Analía AMANDI
In multi-agent environments, autonomous agents need to interact to
achieve their goals because reciprocal dependencies exist among them. In this
context, negotiation is a fundamental tool to reach agreements among agents
with conicting goals. The essence of a negotiation process is the exchange of
proposals. Agents make and respond to proposals in order to converge towards
a mutually acceptable agreement. However, not all approaches are restricted
to that exchange. Several approaches to automated negotiation have been de-
veloped. One of them is the argumentation-based approach [15, 26, 24, 22, 5, 9].
In argumentation-based approaches, agents are allowed to exchange some addi-
tional information as arguments, besides the information uttered on the proposal
[24]. Thus, in the negotiation context, an argument is seen as a piece of informa-
tion that supports a proposal and allows an agent either (a) to justify its position
in the negotiation, or (b) to inuence other agents' position in the negotiation
[12].
When a conict arises during the negotiation, an agent must observe
the negotiation context and determine what arguments can be uttered in order
to reach an agreement. At this point, argument generation can be carried out
in several ways. One of them involves using explicit rules [15, 24]. A rule for
argument generation establishes a set of conditions that the negotiation context
must full to generate a given argument. For instance, if we want to generate a
reward, we need to know what the target of the argument would like to receive
in exchange for its proposal acceptance. In formal terms, to generate a reward,
an agent
a
i
, which needs to persuade an opponent
a
j
, has to observe in the
negotiation context a goal
g
j1
that must be achieved by
a
j
and an action
t
A
that produces the fullment of such goal
g
j1
. Thus, if the agent nds these
facts in the negotiation context, it can generate a reward by saying: if you (
a
j
)
accept my proposal, I promise you to perform action
t
A
.
In most argumentation-based negotiation frameworks, the rules for argu-
ment generation are dened in design time. However, no techniques are dened
to learn from other agents how the agent must build arguments. In fact, people
learn to build arguments from experience. That is, based on the arguments
that a person receives from others and the context in which these arguments are
uttered, he/she is able to infer how the arguments are built and what context
facts are included in the conditions needed to generate those arguments.
In this paper, we propose an approach that allows an agent to learn how
to build arguments by observing how other agents argue in a negotiation context.
Agents that learn how to generate arguments from other agents
3
Specically, our approach focuses on how to learn rules to generate rhetorical
arguments. To infer these rules, we utilise an algorithm for mining fuzzy gen-
eralized association rules [11]. Performing this algorithm, we obtain rules that
associate a set of conditions (antecedents) with an argument (consequent), such
is the format of the argument generation rules that we intend to nd.
Mining association rules allows us to obtain rules from a set of transac-
tions stored in a database. As introduced by Agrawal et al. [2], given a set of
transactions, where each transaction is a set of items, an association rule is an
expression X
Y, where X (antecedent) and Y (consequent) are also sets of
items. That is, the transactions in the database which contain the items of X
will also contain the items of Y. So, if an agent observes and stores in a database
the arguments generated by another agent and the context facts that can be part
of the conditions to generate these arguments, the agent will be able to infer the
rules for argument generation that the other agent applies.
However, the question arises: is it sucient to utilise a traditional (crisp)
algorithm for mining association rules (e.g. Apriori algorithm [2]) to infer argu-
ment generation rules? We have concluded that it is not. There are two factors
that determine this answer: generality and uncertainty.
First, we want the agent to apply the learned rules for argument gener-
ation in dierent negotiation contexts, that is, we need to obtain general rules.
Nevertheless, arguments and facts observed by the agent are expressed in con-
stant terms, because they were uttered in a particular negotiation context. So, if
we utilise a traditional algorithm for mining association rules, the learned rules
will also be expressed in constant terms. For this reason, we opt for an algorithm
for mining generalized association rules. These algorithms use the existent hier-
archical taxonomy of the data to generate dierent association rules at dierent
levels in the taxonomy [28]. Our aim is to build a hierarchical taxonomy of
conditions and arguments in which the leaves are the facts of the negotiation
context (conditions) and arguments observed by the agent, and the upper levels
are the same propositions but expressed in variable terms with dierent degrees
of generality. Then, the generalised association rules algorithm will especially
be able to generate rules at upper levels in the taxonomy of conditions and
arguments. Therefore, the rules will be variable.
The second problematic factor is uncertainty. An agent observing other
agents during the negotiation can only be certain of the arguments uttered,
but cannot be sure of the conditions that these agents check to generate such
4
Ariel MONTESERIN and Analía AMANDI
arguments. The reason behind this fact is that usually it is not necessary to
include all the information used to generate an argument in its premises and
that agents maintain their information private. Despite this uncertainty, the
agent can access to information in the negotiation context that could be part
of the conditions to generate an argument. For instance, following the previous
example, the contextual information around the reward could be: agent
a
j
has
goals
g
j1
,
g
j2
and
g
j3
; agent
a
k
has the goal
g
k1
; agent
a
i
knows that perform-
ing action
t
A
enables it to full goal
g
j1
, that performing action
t
B
enables it
to attain goal
g
j2
, and that performing action
t
A
enables it to full goal
g
k1
.
These facts are present in the negotiation context, but not all this information is
necessary to generate the reward. For this reason, the agent has to dierentiate
relevant from irrelevant information. Thus, taking into account the information
that can be extracted from the argument (e.g.
a
j
is the target, action
t
A
is
the reward), we can determine if a piece of information is semantically related
to the argument. For example, by observing the previous reward, we can see
that the fact that agent
a
i
knows that performing action
t
A
enables it to full
goal
g
j1
is more related to the argument than the fact that agent
a
i
knows that
performing action
t
A
enables it to full goal
g
k1
, since action
t
A
is the action
promised and goal
g
j1
is a goal of agent
a
j
(target of the argument); in contrast,
goal
g
k1
is a goal of another agent not mentioned in the argument.
Therefore, we propose the use of a fuzzy approach for generalised asso-
ciation rules mining to handle this uncertainty. Mining fuzzy association rules
is the discovery of association rules using fuzzy set concepts [17]. The fuzzy set
theory [31] has been used more and more frequently in intelligent systems be-
cause of its simplicity and similarity to human reasoning [13]. Fuzzy sets are sets
whose elements have degrees of membership. In the context of our work, we see
the facts observed in the negotiation context when an argument is generated as
a fuzzy set, where each fact has a degree of membership as regards the semantic
relation between the fact and the argument. Thus, these sets of observed facts
and arguments constitute fuzzy transactions [8]. Consequently, uncertainty is
taken into account during the mining process.
The experimental results showed a high precision of the proposed ap-
proach. To determine the eciency of our approach, we carried out three ex-
periments. First, we compared the rules learned by using our approach with
the original rules used by the observed agents. Second, we compared the rules
learned by using our fuzzy approach with the rules learned by using a crisp
Agents that learn how to generate arguments from other agents
5
one, in order to assess the contribution of fuzziness to this problem. Finally, we
compared the set of arguments that can be generated by using the original rules
with the set of arguments that can be generated by using the rules learned by
the fuzzy approach as well as by the crisp one.
The article is organised as follows. Section 2 introduces basic concepts
about argumentation based negotiation. In Section 3, we present the approach
for learning argument generation rules by observing how other agents argue. In
Section 4, the results extracted from the experiments are presented. Section 5
places this work in the context of previous ones. Finally, in Section 6, we state
our conclusions and suggest future work.
2 Argumentation-based negotiation
In accordance with the work of Rahwan et al. [23], there are two ma-
jor strands in the literature on argumentation-based negotiation: (a) attempts
to adapt dialectical logics for defeasible argumentation by embedding negoti-
ation concepts within them [21, 4]; and (b) attempts to extend bargaining-based
frameworks by allowing agents to exchange rhetorical arguments, such as prom-
ises and threats [15, 24]. Our work belongs to the second strand.
There are several types of rhetorical arguments that an agent can ex-
change during the negotiation. Such types have been commonly studied in the
eld of persuasion in human negotiation [14, 20]. Based on these studies, the
current literature identies at least six types of arguments that an agent can
use during the negotiation [15, 24, 3]. These types are: rewards, used to prom-
ise a future reward; threats, used to warn about negative consequences in case
the counterpart does not accept a proposal; and appeals, used to justify a pro-
posal. Particularly, these appeals can be: appeal to a past reward, to remind
an opponent about a past reward; counterexample, to convey the persuadee a
contradiction between what it says and past actions; appeals to prevailing prac-
tice, to persuade the opponent that a proposal will further its goals since it has
furthered others' goals in the past; and appeal to self-interest, to convince a per-
suadee that accepting a proposal will enable achievement of a goal. In general
terms, a rhetorical argument is composed of four elements: a sender, a receiver,
a conclusion that normally represents the proposal that the argument supports,
and a set of premises that support the conclusion [26].
In an argumentation-based negotiation approach, agents can exchange
arguments in order to justify their proposals, to persuade their opponent, and
