Concurrent unity instances (--num-envs) vs replicated game areas within the scene

Hi:
I have some questions regard using parallelism with mlagents

1. Is there any difference between using concurrent unity instances with the --num-envs argument and replicating several game areas inside the scene? Is one of those options faster? Exists some noticeable difference in the resulting model performance? (assuming equal number of concurrent instances and number of game areas)
2. Is recommendable to use both options at the same time?
3. If it's the case, which one is better?
4. Im using a procedural(stochastic) maze generator in my game. When i train with several concurrent unity instances im getting the same mazes in the same order on each unity instance. Im assuming that each proccess is using the same seed to generate random numbers. How can i change that in order to leverage a better generalization?

Thanks in advance

 

This question has not been answered. I dont know if is not clear enough. Please let me know if that is the case and i will explain better. Thanks.