I'm using DQN to train multi-version of the same system and there is a small difference when I run them both separately. However, my result suddenly dropped in both versions if I run them both at the same time. I tried again but I got the same results with slightly different. Would it be affecting my results if I run multi-version of my system at the same time?

Is there any explanation for that and How can I get accurate results when I train multi-version of the same system at the same time?

  • $\begingroup$ I don't understand what "mult-version of the same system" means. What is different between the versions, and how do the versions interact on the computer system when running at the same time? Do the versions interact because the environment is shared (e.g. you are training a self-driving car of Carla, and putting two DQN agent cars into the same simulated town where they might encounter each other)? Could you please explain what you are doing in more detail in the question? $\endgroup$ Dec 18, 2019 at 15:25
  • $\begingroup$ @NeilSlater Actually, I run the same code with different maps but I run three copies at the same time with different maps but if run one of these three maps alone I got different results. $\endgroup$ Dec 18, 2019 at 15:33
  • $\begingroup$ So there is no intent for the different maps and agents to interact? By "run at the same time" are you using separate processes or separate threads? Are the three agents sharing a GPU for any calculations? Have you tried running one agent multiple times on its own to observe any natural variance during learning? $\endgroup$ Dec 18, 2019 at 15:36
  • $\begingroup$ I run all of them using subprocesses and I don't think that there is sharing of the GPU. There is a small difference when I run them both separately but there is a big difference when I run all of them at the same time. I need to do it this way becuesue it takes 3 days' training. $\endgroup$ Dec 18, 2019 at 15:41
  • 1
    $\begingroup$ OK, I don't think you will get an answer here. The problem is going to be specific to whatever system, libraries and code you are using. I would classify it as a fault/bug, but cannot really say more than that on the information given. $\endgroup$ Dec 18, 2019 at 16:05


You must log in to answer this question.

Browse other questions tagged .