Adaptação e análise do critério de reciprocidade para estratégias de argumentação em diálogos persuasivos com agentes auto-interessados

Abstract

In a persuasion dialogue between two agents, it is possible to use rhetorical arguments, which include threats and rewards. Threats are persuasive arguments, but they reduce the opposing agent’s utility, while a reward does not reduce your adversary’s utility, but is less persuasive. In an environment with self-interested agents, the tendency is for agents to primarily use threat arguments, since they are not interested in the loss of utility of their adversaries. However, this implies a loss of general utility for all agents, as all agents will be adopting a more threatening strategy to try to persuade their adversaries. In order to try to solve this problem, this work seeks to adapt a criterion for strategies applied in game theory that reduces the total utility loss of agents, even if they are self-interested. This criterion is called reciprocity, and its main characteristic is to always reciprocate an action taken by your adversary. In our case, if an adversary threatens an agent that has reciprocity, he will threaten back, while an adversary that rewards an agent that has reciprocity will be rewarded back. Despite being simple, strategies that adopt this criterion are widely used and are usually the winners when placed to compete against other strategies. To analyze the persuasiveness and threat reduction capacity of the proposed strategy, simulations of persuasive dialogues with rhetorical arguments were performed, and within these tests, experiments were also carried out to analyze the bias of the test parameters established for the generation of scenarios. The experiments carried out showed an increase in the agents’ usefulness when using the strategy with the reciprocity criterion, reaching the research objective of this work. The main contribution of this work is in adapting reciprocity from game theory to argumentation systems and understand the advantages and disadvantages of basing a strategy on this criteria with an empirical test. As a secondary contribution, this work also proposes an adaptation of the Dung preferred argumentation semantic that chooses only the arguments from one of the agents to be applied in persuasive argumentation.