I designed a DQN architecture for some problem. The problem has a parameter $m$ as the number of clients. In my situation, $m$ is large, $m\in\{100,200,\ldots,1000\}$. For this situation, the number of input ports of the DQN is some few thousand, $\{1000, 2000, \ldots, 10000\}$. For some fixed $m$, I would like to see the performance of deep Q learning on the performance. So I have to train the DQN for every change that occurs on $m$ and this should handle thousands of inputs ports for each training. Is this situation familiar in DQN and if not how to solve this issue?

  • 1
    $\begingroup$ By input ports are you talking about the state space being an $m$ dimensional vector, or the action space? $\endgroup$ Aug 31 '20 at 17:56
  • $\begingroup$ Yes it is the state space, and it is a $O(m^2)$-dimensional vector. $\endgroup$
    – zdm
    Aug 31 '20 at 18:47
  • 1
    $\begingroup$ that is...large. I've never experimented with a state space close to that size. I would imagine you'd first need to have a network capable of handling such large dimensional inputs and even then I'm not sure how feasible it would be to train an agent given the amount of exploration that would need to be done for a state space that large. $\endgroup$ Aug 31 '20 at 18:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.