A stability-based approach to inquiry dialogues in agent argumentation

Abstract

In inquiry dialogues, two or more agents work together to prove or disprove a proposition. The de facto standard for this type of dialogues are exhaustive systems, in which agents make every move that could impact the outcome of the dialogue before terminating the dialogue. This behaviour results in positive properties such as soundness and completeness, but as a downside generates long dialogues including redundant moves. New research proposes a stability-based querying system, in which agents stop making queries when it is certain that the outcome of the process will not change anymore. In this thesis, this querying system is extended into a set of three, first of their kind, stability-based inquiry dialogue systems with increasing levels of expressiveness. First it is explored how performance of inquiry systems can be defined and compared. Next, the performance of these stability-based systems is compared to each other and to exhaustive systems. Experiments are performed on multiple rule sets in which the agents try to minimize either the dialogue length or the amount of observations shared. Training and test sets for these experiments are initiated using three different instantiation functions to model different cases that could occur in real life situations. It is shown that stability-based systems in general outperform the exhaustive systems and that, depending on the performance measure and instantiation function used, the additional expressiveness results in increased performance as well. The downside of the additional expressiveness is an increase in complexity, resulting in the more expressive systems not being able to perform well on large problems.