Let's Play Again: Variability of Deep Reinforcement Learning Agents in Atari Environments