• PhD Student:
  • Quentin Elsaesser
  • Co-Supervisor :
  • Patricia Everaere (CRIStAL)
  • Funding : ANR
  • PhD defended on :
  • Dec 16, 2024

For both humans and artificial agents, one of the main ways to gain information and knowledge about our environment is to interact with other agents or sources of information. These agents or sources will send us information that we can then take into account to form our own opinions. This information is potentially conflicting. It is therefore necessary to be able to form an opinion and to evaluate the reliability of these agents and of the information received.

This thesis proposes a family of methods for evaluating the reliability of agents and information. We have no a priori knowledge of either the agents or the information collected. Therefore, we evaluate reliability solely by comparing the different opinions of the agents. We are interested in epistemic questions and our aim will be to search for the truth. This search for truth will be based on the idea of Condorcet’s Jury Theorem, which states that it is more likely that the majority of sources will find the correct solution to the questions asked. We study the different properties of our methods and experimentally compare our methods with those in the literature for truth-tracking. We study the different properties of our methods and experimentally compare our methods with those in the truth-finding literature. We propose an evaluation on real data as well as on experimental data, which allows us to cover as many different cases as possible.

We then propose to use propositional formulae to represent our information. To do this, we will focus on two areas: judgment aggregation and belief fusion. The goal is to make the best decision by taking into account the opinions of the agents.

We propose a new formula-based belief fusion operators that evaluates the reliability of agents and then makes a decision based on this reliability. We then adopt an epistemic view of belief merging, then the purpose of the merging process is to approximate the true state of the world. We also examine the IC postulates satisfied by our new operator and propose an experimental evaluation showing that reliability still produces better decisions than directly using the number of votes obtained by a formula.

Finally, we propose a new family of methods for aggregating judgments. We show that using reliability rather than the number of votes gives better results for the best possible decision, especially in more complicated cases such as when the number of agents is even and there are ties in the number of votes obtained by the formulas. We also see which standard judgment aggregation properties are satisfied by our methods.

Committee

Reviewers

  • Andreas Herzig, directeur de recherche CNRS, IRIT, Toulouse
  • Nicolas Maudet, professeur des universités, LIP6, Sorbonne Université

Examiners

  • Sylvain Lagrue, professeur des universités, Université de Technologie de Compiègne
  • Pierre Marquis, professeur des universités, CRIL, Université d’Artois

Supervision

  • Patricia Everaere, maître de conférences, Université de Lille
  • Sébastien Konieczny, directeur de recherche CNRS, CRIL, Lens