Hyperparameters tuning for deep reinforcement learning

Hello,

I have a deep reinforcement learning agent that takes 2 hours for training (exploration). How to tune its hyperparameters (learning rate, number of hidden layers, number of neurons …etc) when we need to test many reward function designs?

I think that hyperparameters tuning will hard regarding the infinite combination for the hyperparameters values.

Hi @ZINE ,
In Week 3 of course 2 you have tree complete videos dedicated just to Hyper-parameter Tuning, but to give a brief introduction:

  • Not all hyper-parameters have the same importance, so you can start playing with the most important first (this is described in the first video).
  • Also, you have to carefully choose the range of values each hyper-parameter can take (video 2)
  • Tuning will take time, so there are different strategies to use, depending in the computing capabilities you have at your disposal (video 3).

As a side note, most cloud AI providers perform the tuning for you, like Amazon , Azure and Google.

1 Like