Description
Date depot: 24 septembre 2021
Titre: Leveraging Self-Play to Improve Task-Level Generalization in Reinforcement Learning
Directeur de thèse:
Olivier SIGAUD (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Reinforcement learning, coupled with self-play, has demonstrated its ability to create
agents able to solve complex sequential decision-making problems such as playing Chess,
Go, StarCraft and Dota. However, these agents are limited. While being expert in their
respective task, these AI systems are not able to deal with other tasks, even if those tasks
share some structure. This motivates us to study the following question, which will be the
focus of this thesis. How can we leverage self-play to build more generally capable
agents, able to solve a vast variety of tasks? Even though the thesis will be more
oriented toward experimental than theoretical research, theory will remain part of it as it
helps to better grasp underlying concepts and consequently help with experiments.
Doctorant.e: Cachet Théo