Reducing the Computational Cost of Deep Reinforcement Learning Research