Bounded Ratio Reinforcement Learning