A0796
Title: Using Bayesian neural network as an actor in actor-critic methods
Authors: Leo Grill - Universite de Poitiers (France) [presenting]
Yousri Slaoui - University of Poitiers (France)
David Nortershauser - Orange (France)
Stephane Le Masson - Orange (France)
Abstract: Reinforcement learning and deep learning lack statistical theory background. Bayesian approaches lead to more robust learning. The Bayesian theory has already started enlightening Deep learning. It can be used to deal with issues such as overfitting or black-box modeling. Bayesian methods are used to train the neural network encoding the actor and find a policy in deep actor-critic algorithms. The results of simulations are presented to show the interest in using these methods for deep reinforcement learning. It is particularly important for the exploration phase and for the stability of learning. The Bayesian neural network also benefits from the advantages of ensemble methods, it can be used to realize coherent exploration. The prior and posterior approaches can be either used to regularize the model or two bring some knowledge during the training.