Description
Date depot: 12 avril 2022
Titre: Social learning and minimal models of cognition in swarm robotics
Directeur de thèse:
Nicolas BREDECHE (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: cf. document ci-joint pour une version lisible // see PDF for readable version ---
// commission IA-BD //
Mots clés: apprentissage social, cognition sociale, robotique en essaim, robotique évolutionniste. //
Keywords: social learning, social cognition, swarm robotics, evolutionary robotics. //
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 decentralised manner to novel environmental conditions. Individuals in these systems must be able to acquire new behaviours autonomously (through natural evolution or social learning), and they must be able to do so in an open-ended manner, to deal with potentially entirely novel environments.
In this thesis, we will study how to implement social learning algorithms on swarms of autonomous robots with limited computation and communication capabilities. These social algorithms rely on two complementary processes: (1) innovation through individual learning and (2) diffusion of successful behavioural strategies throughout the swarm. Social learning is then a continuous learning process that is used to learn and share efficient behavioural strategies, even when the environment is non-stationary, which will be the context of interest for this thesis.
Social learning in swarm robotics is currently limited in two ways. Firstly, learning is considered to act at the level of the policies' control parameters, such as the weights of the artificial neural networks used to map sensory inputs to motor outputs. Secondly, the type of policies that are currently under scrutiny is either reactive policies that make the optimistic assumption that all information to make the right decision is contained in the current sensory inputs or policies with limited memory use which is usually modelled by recurrent neural networks with few nodes.
This thesis will address these two limitations. First, we will consider how control policies may be improved by observation rather than by directly copying control parameter values. This is actually a topic of growing interest in the reinforcement learning community in a teacher-observer setting, but which we will address in the context of many (>100) robots. Secondly, we will consider what are the low-level cognitive capabilities, such as the ability to maintain social recognition of potential partners and reliability estimation, that can give rise to more efficient self-organization behaviours in a robot swarm through cooperation. As such, the expected results from this thesis are a new class of social learning algorithms through observation and social cognition, which is primarily targeted at the domain of learning in swarm robotics, but we may also provide modelling tools for a better understanding of social learning and cognition in animals.
Finally, we aim at experimenting both in simulation and with real robots (if relevant), using a swarm of robots already available at the Institut des Systèmes Intelligents et de Robotique, Sorbonne Université.
Cf. fichier PDF attaché pour les références bibliographiques.