Stopping conditions for DQN training

Hello all,

I am currently playing around with DQN trainning. I am trying to find a systemic way to stop the trainning process rather than to stop it mannually. However, for my trainning process, I have no idea what the end rewards will be and I don't have a target point to reach. Therefore, I do not know when to stop.

Is there a way for me to stop DQN agent without those information and guarentee some type of convergence?

Thanks for helping!

Answer 1

[Madhav Thakker]

Hi Zonghao zou,

One possible parameter to consider when stopping training is Q-Values. If the Q-Values are saturated, it means that no learning is happening in the network. You can perhaps look at your Q-values and decide a threshold, to perform early-stopping in the network. You don't need the end-reward or target-point to perform early stopping based on Q-Values.

Hope thi helps.