I'm trying to implement an agent in RL and I was wondering if it's pertinent to take random actions in DeepRL. I see that in traditional Q-learning that sometimes we take random actions to encourage exploration (and gradually reduce some epsilon to reduce the exploration)

I did some googling about it being used in DRL, and went through some medium posts but no one seems to talk about it. The fact that it's not brought up (in the posts or code) makes me think that this is no longer a thing. I did find some sources Qlearning Epsilon-greedy exploration: Epsilon decay X fixed and Epsilon and learning rate decay in epsilon greedy q learning and I'm not sure if they are pertinent to deep RL. I have also looked through Exploration in Q learning: Epsilon greedy vs Exploration function and DQN fails to find optimal policy

Do I need to be sampling based on the current agent or can I just use any arbitrary sampling method? And is epsilon sampling of random actions actually useful in Deep RL or is it more of a thing in standard RL?


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.