Improving Generalization in Game Agents with Imitation Learning