Cálculo da força de argumentos retóricos e sua utilização em diálogos de negociação persuasiva em sistemas multiagente

Abstract

A negotiation between agents is called persuasive when the proposals are backed by rhetorical arguments (threats, rewards, or appeals), whose role is to try to persuade the opponent agent to accept the proposal more readily. This thesis tackles the problem of calculating the strength value of these kinds of arguments. In the related work, the strength value of a rhetorical argument is represented by a vector of two elements: the value of the uncertainty of the beliefs that make up the argument and the value of the importance of the opponent’s goal. Nevertheless, there is a need of a further analysis of these components and of the characteristics of the participant agents that may impact on the strength value. Therefore, the objective of this work is to study these kinds of arguments and to propose a more expressive model for calculating their strength values. This thesis contains three main parts. The first one concerns the design of an agent architecture that is based on the goal processing model proposed by Castelfranchi and Paglieri, which can be considered an extension of the Beliefs-Desires-Intentions (BDI) model. In this model, the goals go through four stages of filtering from being mere desires until they become an intention. We propose an argumentation-based formalization of this model, which means that the passage of the goals from one stage to the next is supported by arguments. The second part of this thesis concerns the strength value calculation model. First of all, the architecture of the negotiation agents and the logical definitions of each kind of rhetorical argument are presented. After that, the criteria that are taken into account for the strength calculation are presented. Thus, besides considering the importance of the opponent’s goal, we also consider the effectiveness of this goal and the credibility of the participant agents. The effectiveness of the opponent’s goal is calculated based on its status – according to the model of Castelfranchi and Paglieri – and the kind of rhetorical argument it makes up. The last part presents a set of experiments that aims to empirically evaluate the proposed model. With this purpose, firstly, a negotiation model that rules the behavior of the participant agents during the dialogue is presented. The experiments evaluate the efficiency of the proposal by comparing it with the closest proposal found in literature. The results demonstrate that the proposed model is more efficient in terms of number of negotiation cycles, number of exchanged arguments during the negotiation, and the number of achieved agreements.