Correlation Priors for Reinforcement Learning