I'm looking for some indications about the tuning of hyperparameters in building my double DQN. I have a time series problem (with about 2000 observations and no terminal state, I have to max the rewads) so I decided to define as an episode a subsequence of N observations (from that dataset) starting from a random point (e.g. from the 245th period to the 245+N period). So here the first doubt: should I run more episodes with a smaller N or vice versa? In a different context I would make some tries, but here it takes a lot of time.

What about the length of the batch and the step when updating? Are there any generally accepted rules?

The final hyperparameter that I don't know how to choose is the moment when updating the target model weights. I know I should do that rarely, but I do not find any specific rule (if not very high threshold, that I doubt in my problem will be reached).


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.