Watch your Weight Reinforcement Learning

Deep Neural Networks (DNN) have been a crucial driver behind the recent successes of deep reinforcement learning (RL). While allowing for a tremendous scale-up to high dimensional domains like robot-control or GO they have also introduced various difficulties into the training process, that have been addressed by methods like double Qlearning to stabilize learning. However, the role of the weights is normally limited to a storage object. This work connects the information in the weights via a PAC-Bayes bound to generalisation across a distribution of environments and showcases maximum-entropy and exploration via noise injection as places to exploit the connection of information in weights and activations.