Speedy q-learning: a computationally efficient reinforcement learning algorithm with a near optimal rate of convergence