Probing various layers of DQN to assess "difficulty" of decisions. Focusing on model trained on the atari asteroids game.