Fiche de FONTBONNE Nicolas

Titre : Individual and group learning dynamics in evolutionary collective robotics
Date : 14 mars 2023
Heure: 17:00
Résumé : This thesis discusses the problem of coordination between robots in situations where they must learn to interact with each other autonomously. We propose two machine-learning algorithms to address this problem. The first algorithm, Horizontal Information Transfer (HIT), is a networked evolutionary learning algorithm that allows robots to learn their control policies online while interacting with their environment and exchanging parameters. We study the dynamics of this algorithm on simulated foraging tasks. The second algorithm, called CC-1+1-GA, is a Centralized Cooperative Co-Evolutionary Algorithm (CCEA) that modulates the number of agents' policies modified at each generation to find a balance between individual and group learning. This algorithm is tested on a resource selection tasks and a simulated multi-rover exploration problem. We show that these methods handle multi-agent exploration noise issues while allowing behaviours to be coupled during the learning process.
Lieu : ISIR, Tour 65/66, Étage 3, Salle 310,