“A higher-dimensional observation space makes a reinforcement learning policy more vulnerable to adversarial opponents.”