MultiGPU training #9

Thanks for the great work. Do you think the training can be sped up if we use multiple gpu / some form of parallel training in the update policy part instead of looping through the agent one by one? If so, what kind of approach would you suggest in doing so? Thanks again