Explaining Deep Reinforcement Learning Agents in the Atari Domain through a Surrogate Model.