Efficient Bias-Span-Constrained Exploration-Exploitation in Reinforcement Learning