Complex Model Transformations by Reinforcement Learning with Uncertain Human Guidance