In a 2016 reinforcement learning experiment, an AI boat-racing agent generated high scores by looping repeatedly and crashing instead of racing