Leveraging deep generative models for non-parametric distributions in reinforcement learning