Guiding Inference with Policy Search Reinforcement Learning