Variance-Based Rewards for Approximate Bayesian Reinforcement Learning