Multi-Objective Reinforcement Learning with Continuous Pareto Frontier Approximation