Knowledge Evolution in Agent Populations

par Line Van Den Berg

Projet de thèse en Mathématiques et Informatique

Sous la direction de Jérôme Euzenat et de Manuel Arcas.

Thèses en préparation à Grenoble Alpes , dans le cadre de École doctorale mathématiques, sciences et technologies de l'information, informatique (Grenoble) , en partenariat avec Laboratoire d'Informatique de Grenoble (laboratoire) et de EXMO - Echanges de connaissance structurée médiatisées par ordinateur. (equipe de recherche) depuis le 01-10-2018 .

  • Titre traduit

    Knowledge Evolution in Agent Populations

  • Résumé

    When two populations of agents encounter, they do not necessarily organise their knowledge about their environment in the same way. They may however attempt at communicating and progressively align their knowledge. We aim at studying the effectiveness and robustness of such a process. These problems may be approached either theoretically or experimentally, through the framework of cultural evolution. Experimental cultural evolution provides a population of agents with interaction games that are played randomly. In reaction to the outcome of such games, agents adapt their knowledge. It is possible to test hypotheses by precisely crafting the rules used by agents in games and observing the consequences or to prove analytically properties of such settings. Our ambition is to adapt the successful cultural language evolution approach [Steels, 2012] to the evolution of the way agents represent knowledge [Euzenat, 2014; Anslow & Rovatsos, 2015; Chocron & Schorlemmer, 2016]. We have applied this approach to ontology alignment repair, i.e., the improvement of incorrect alignments [Euzenat, 2014; 2017]. For that purpose, we performed a series of experiments in which agents react to mistakes in ontology alignments —expressing relations across ontology concepts [Euzenat & Shvaiko, 2013]. Agents only know about their ontologies and alignments with others and they act in a fully decentralised way. We showed that cultural repair is able to converge towards successful communication through improving the objective correctness of alignments. This PhD proposal focusses on the behaviour of populations of agents as opposed to agents individually. Agents may belong to different populations, though still playing individual games. There are two important questions when doing so: What characterises a population? It can be geography, interbreeding, the topology of agent connections, the shared culture (here knowledge) or the capability to communicate. How the knowledge of an agents impacts that of the population it belongs to? It may be that among the same populations agents have a better communication, hence knowledge propagate more easily (this comes for free if the population is based on a topological basis and agents communicate directly), or that they have specific ways to share knowledge, e.g., synchronisation (called alignment in [Steels, 2012]). So our goal is to study how the answer to these two questions impact the propagation/evolution of knowledge among agents. This may be developed, for instance, through extending the alignment repair games that we developed: agents in a population, sharing the same ontology, can take advantage of what is learnt by the others agents interacting with agents of another population. Games will have to be designed for agents to locally adapt alignments between their ontology and those used by other populations. It is expected that different agents, in the same population, do not necessarily end up with the same alignments. We want first to understand when this occurs as well as what can be done for the agents to share with their peers the correspondences that they found. Roughly, three modalities can be compared: Agents develop independent alignments which can be compared; Agents develop independent alignments, but from time to time, proceed to synchronisation when they exchange their findings (this synchronisation, in turn, may be global to the population or local to a pair of agents); Agents develop from the beginning shared alignments. In the two latter modalities, different operations may be used to aggregate the results brought by other agents. Finally, it is possible to consider more than two populations and/or ontologies, eventually by splitting and merging populations and to study its impact on the alignment process. This raises the problem of the evolution of populations especially when what makes a population is characterised by its shared knowledge. There is then co-evolution of knowledge and populations, as was already observed in [Axelrod, 1997]. This work is part of an ambitious program towards what we call cultural knowledge evolution. Its results may be of experimental or theoretical nature and it may provide practical, e.g., new adaptation operators, or methodological, e.g., better experimental procedures, contributions.