Pre-training of Deep RL Agents for Improved Learning under Domain Randomization