Reinforcement Learning with Generative Models for Compact Support Sets