On the Optimality of General Reinforcement Learners