The adaptive k-meteorologists problem and its application to structure learning and feature selection in reinforcement learning