Description
Date depot: 1 janvier 1900
Titre: Evolution of cooperation in collective adaptive systems (including swarm robotics)
Directeur de thèse:
Nicolas BREDECHE (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
Advisor #1 / directeur: Nicolas Bredeche, ISIR/UPMC (nicolas.bredeche@isir.upmc.fr)
Advisor #2 / co-directeur: Jean-Baptiste André, ENS (jeanbaptisteandre@gmail.com)
Mots clés: évolution de la coopération, robotique en essaim, robotique évolutionniste, coopération mutualiste, choix du partenaire
Keywords: evolution of cooperation, swarm robotics, evolutionary robotics, mutualistic cooperation, partner choice
An important challenge in collective systems, whether natural (e.g.: group of hunters) or artificial (e.g.: swarm robotics), is to understand how they are able to adapt in a decentralized manner to novel environmental conditions. Individuals in these systems must be able to acquire new behaviors autonomously (through natural evolution or learning), and they must be able to do so in an open-ended manner, to deal with potentially entirely novel environments.
However, decentralized adaptation poses an important scientific problem as nothing guarantees that adaptive mechanisms taking place at the individual level lead to the establishment of behaviors that are adaptive at the collective level: collective adaptation does not always follow from individual adaptation. First of all, many collectively efficient behaviors require the coordinated action of several individuals, and cannot be reached by the independent improvement of each individual separately. Second, and worst, individual adaptations can even harm, rather than help, collective efficiency if individuals have externalities for each other (e.g. if they enter into competition with one another, if they can underinvest into a collective good, etc.), in which case decentralized individual adaptation shall lead to a reduction of performance of the entire system.
Ever since Darwin, evolutionary biologists have acknowledged and studied this problem [1]. They have proposed various mechanisms through which collectively efficient outcomes (so-called cooperative outcomes) can be reached via individual adaptation [2-6]. Identifying the conditions in which individual adaptation can, or cannot, generate collectively efficient outcomes is indeed important to understand the origin of cooperation in biology, but it is also key to the design of practical solutions for open-ended adaptation in collective artificial systems such as swarm robotics [7,8].
The objective of this thesis will be to explore (at least) one important potential solution to this problem known from evolutionary biology: the role of reputation [9-11]. We will explore whether socially efficient outcomes are more easily reached by individual adaptation, when individuals can recognize others and are informed of their past behavior (their reputation). To this end, we will use evolutionary robotics [12], that is: the artificial evolution of embodied agents, in order to provide an accurate simulation of interactions between individuals.
Evolutionary robotics has originally been proposed as an optimisation method to automate the design process of the decision-making control architectures for complex problems, including those met in collective robotics. It also provides a very convenient framework to extend classical models used in evolutionary game theory with the ability to model and simulate the mechanistic aspects underlying cooperation, such as physical coordination between individuals.
Doctorant.e: Ecoffet Paul