I was wondering about hyperparameters used in Deep Q-Networks. Considering the use of replay memory and target network, together with the epsilon-greedy policy, are the number of epochs different of 1 useful? In the affirmative case, which are the situations and benefits?

The reason is that we want to explore the environment, thus providing a high number of different samples, but at the same time, my intuition says that it can be useful for fast convergence of the Q-values.

  • $\begingroup$ Could you clarify what you mean by "are the number of epochs different of 1 useful?" $\endgroup$ Commented Oct 21, 2020 at 4:43
  • $\begingroup$ It's simply the fact that epochs are > 1. $\endgroup$
    – HenDoNR
    Commented Oct 21, 2020 at 6:45


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.