System Identification meets Reinforcement Learning: probabilistic dynamics for regularization

Model-based and model-free perspectives are two well established paradigms in RL. in this thesis a mixed approach is proposed, in which the interactions with the real system are carried out in both ways: a rough model is retrieved in order to play the role of a regularizer, while the punctual estimation over specific values of the policy parameter is placing reliable punctual estimates that should be fitted by the reconstructed function.