Logarithmic Regret of Exploration in Average Reward Markov Decision Processes