Max Step in tensorboard doesnt match Max Step of agent

Hi all, i find it strange that Episode Length seems to cap at 4000 steps but my agent’s max step settings is 20 000. Is this a bug?

7398728--903929--upload_2021-8-8_17-54-1.png
7398728--903926--upload_2021-8-8_17-53-12.png

ML-Agents 2.1.0

Max step on the agent ml-agents/docs/Learning-Environment-Design-Agents.md at mapoca-all-envs · Unity-Technologies/ml-agents · GitHub and max step on the training config file ml-agents/docs/Training-Configuration-File.md at mapoca-all-envs · Unity-Technologies/ml-agents · GitHub mean two different things. They are named the same due to unfortunate circumstances.

@vincentpierre yes i understand that. correct me if i’m wrong, max step on the agent corresponds to the maximum episode length? if so, i’ve set it to 20 000, but episode length is ending at 4000 regardless.

Ah! Sorry about that. I believe it is because the steps of the agent are counted in FixedUpdates and not in number of decisions. On the other hand, the episode length in tensorboard is reported in number of decisions. I think the issue you have is because you have a decision requester with a decision period of 5. If you want 20000 decisions, you need to set the max step to 20000 * decision period = 100000. The reason the code is like this is because the decision requester is not a required component for training, so we cannot use decision period at all when reporting the episode length or counting steps.
I hope this helps.