Description du projet
L’apprentissage par renforcement fait un grand bond en avant
L’apprentissage par l’interaction est une idée fondamentale sous-jacente à presque toutes les théories d’apprentissage et à l’intelligence. Comparé à d’autres approches d’apprentissage automatique, l’apprentissage par renforcement est davantage centré sur les problèmes de calcul qui se posent lorsque les agents logiciels apprennent en se basant sur les interactions avec un environnement. Le projet COLT-MDP, financé par l’UE, entend faire progresser les approches théoriques de pointe de l’apprentissage par renforcement. L’étude s’articulera autour de trois piliers. La conception de modèles compacts de représentation permettra aux chercheurs de mieux structurer des problèmes complexes. L’apprentissage par renforcement pourrait alors être étendu à de nombreux domaines. En outre, les chercheurs développeront des algorithmes de calcul efficaces. Inspiré par les préoccupations du public, un organisme croissant de recherche abordera les questions liées à l’équité et à la vie privée dans l’intelligence artificielle.
Objectif
Computational learning theory has been highly successful over the last three decades, both in providing deep mathematical theories and in influencing the practice of machine learning. One of the great recent successes of computational learning theory has been the study of online learning and multi-arm bandits. This line of research has been highly successful, both theoretically and practically, addressing many important applications. Unfortunately, the recent theoretical progress in Markov Decision Process and reinforcement learning has been slower.
Based on my fundamental contributions to reinforcement learning (e.g. policy gradient, sparse sampling and trajectory trees), to online learning and machine learning in general, I propose to take the theoretical and practical success of online learning to the “next level” by making significant breakthroughs in reinforcement learning. Our main aim is to advance the state of the art in the theory of reinforcement learning, and our research will revolve around three pillars: (1) compact representation, (2) efficient computation and (3) societal challenges, including fairness and privacy.
A successful project will greatly impact reinforcement learning in all its stages. Modelling: Introducing new compact representation models, will enhance our understanding how to structure complex problems, which would greatly extend the applicability of reinforcement learning. Efficient computation: New algorithmic methodologies will give new insight for overcoming computational and statistical barriers both for planning and learning. Learning: New learning paradigms would address fundamental issues of copping with uncertainties in complex control environments of reinforcement learning. Societal challenges: Allowing the community to understand, assess, address and overcome societal challenges is of the greatest importance to the acceptance of AI methodologies by the general public.
Champ scientifique
Programme(s)
Thème(s)
Régime de financement
ERC-ADG - Advanced GrantInstitution d’accueil
69978 Tel Aviv
Israël