Discovery of Optimal Solution Horizons in Non-Stationary Markov Decision Processes with Unbounded Rewards