Exécution de notre outil de synthèse sur un nouvel article

Résumé : Offline reinforcement learning promises to enable control policy learning directly from previously collected data. It is a challenging problem because data collected from previous policies may not well-represent the distribution of the data when following the learned policy. This work evaluates the performance of recent offline RL algorithms using the library D3RLpy, which includes recent state-of-the-art algorithms as well as relevant datasets to evaluate them. We evaluate the performance of algorithms provided by the library without hyperparameter tuning. We compare performance both in terms of algorithm performance on the environments, as well as considering the effect of the dataset on the performance a few select algorithms.