Reinforcement Learning Toolbox - When does algorithm train?

I am currently using the RL-Toolbox with a DQN-Agent built into a long-running process-simulation.

The maximum stepcount is currently 8000 steps per episode.

Unfortunately the documentation seems a little ambiguous to me, so here my question:

Doese the train-function of the RL-Toolbox train the agent at the end of an episode or during the episode when the step count exeeds the minibatch-size (like in the baseline algorithms)?

Thank you in advance.

Answer 1

[Emmanouil Tzorakoleftherakis]

The implementation is based on the algorithm listed here.

Weights are being updated at each time step.

[Hans-Joachim Steinort]

"For each training time step" - that was the line I was looking for (yet looking into the source code lead me to the same conclusion).

After double-checking the baseline-algorithms I found that they do it the same way.

Thank you for your time!