Reinforcement Learning with Attention that Works: A Self-Supervised Approach https://arxiv.org/abs/1904.03367