DQN learns at first but then worsens.

Hi, I am training a DQN agent with a simevent model. I am testing out different hyperparameters, but everytime the agent learns (reward goes higher) at first for a while, but then goes down. I have tested different learning rate, exploration epsilon, and discount factors. But the shape of training progress is pretty much same in all combinations. Is there any potential way I can fix this issue?

Answer 1

[Emmanouil Tzorakoleftherakis]

To confirm that this is an exploration issue, can you try setting the EpsilonMin param to a high value? e.g. 0.99. If after doing that you still see the same result, there is likely something else going on.

[Khandakar Rashid]

Thank you Emmanouil for the suggestion. I have tried Epsilon = 1, EpsilonMin=0.99. Unfortunately, no luck :(

Do you have any other tips?

[Emmanouil Tzorakoleftherakis]

Hard to tell, but it's strange to me that the episode curve is similar every time. That makes me think that there is something specific about the way you have modeled your environment model that guides the training through a similar path each time.