Tech Reports

ULCS-10-003

Arguing from Experience: Persuasive Dialogue based on Association Rules (PhD Thesis)

Maya Wardeh

Abstract

The development of autonomous software agents requires consideration of a number of elements. One interesting aspect of the study of software agency is to enable effective inductive reasoning, amongst agents, using accumulated experience of an agent. However, this experience may vary from agent to another, and only by exploiting these differences, in the right way, can the agents come to an agreement regarding some issue. This thesis is concerned with one particular aspect of such agency: modelling the process of arguing from "experience" to equip autonomous agents (entities) with the capability to jointly coming to a "view" regarding some case, using the experience they have independently gathered over time. The background setting for this work deals with the topic of induction as a dialectical form of reasoning, and attempts to address some issues regarding its treatment in philosophy, as well as the problems inherent in the computational modelling of such reasoning. The main output of the study is a model to enable "Arguing from Experience" which uses techniques from the field of argumentation theory and knowledge discovery in databases, to enable agents to pool and construct arguments in support of and against proposals for "views" regarding some given case. Arguments are pooled from the agent's experience by means of Association Rule Mining (ARM) techniques. The proposed model is intended to describe how "Arguments from Experience" can be put forward and be systematically attacked in a variety of ways. This enables the agents to consider all available options and come to a conclusion about the best "view" to associate with the given case. The underlying theory extends a well established account from the field of philosophy, based on the use of argument schemes and critical questions. The account given is then formalised in terms to enable its representation in agent systems.

The underlying model has formed the basis for two applications: an implementation of a dialogue game protocol to provide a proof of concept, as well as means to enable automated two-party "Arguing from Experience" for the purposes of classification; and a framework to aid multiparty "Arguing from Experience", intended as means to facilitate argumentation between more than two parties. Both applications are evaluated using variety of setups in various domains. The obtained results provided empirical evidence to the efficacy of "Arguing from Experience".

For each technical report listed here, copyright and all intellectual property rights remain with the respective authors. Copyright is effective from the year of publication in each case. By downloading a file from this page, you agree to use it only for purposes of research and scholarship. Any other use of this material or storage of it in any medium or its sale or distribution in any form is expressly forbidden without prior written permission from the authors concerned.