What should I do when I want to continue the training ?

[Morgan-6Freedom]

Hello.

I have trained my AI 50k steps (the default max-steps) with this command:
mlagents-learn config/config.yaml --run-id=first-run --train

And I think it hasn't learn yet. Should I relaunch the training with the same command ? Should I only change the "run-id" ? Should I erase everything and modify the max-steps to the number of steps I think is enough ?

Thank you

[MarkTension]

If it's learned anything then you'd probably want to add --load to that run-id to continue where you ended, and increase max-steps.

You say you think it hasn't learned yet after 50k steps? Maybe you want to check if it learned anything at all by checking the progress on tensorboard. If it learned to increase its cumulative reward to some degree then you might want to continue training. If nothing at all happened after 50k steps then you might have some error and you'd probably have to make some changes to your hyperparameter settings or environment if it's a custom one.

[Morgan-6Freedom]

Thank you. I tried to add --load on my second training. But in tensorboard, the graph restart from the left so it's really messy to read anything

[MarkTension]

yeah that's pretty messy isn't it? It also messes up the relative time. You'll get all the info, but i've also not found a way to deal with that yet

[dlindmark]

If I recall correctly, the tensorboard graph continued for mlagents version < 0.13.0. I prefered it that way, since it makes it for cleaner graphs.

[stale]

This issue has been automatically marked as stale because it has not had activity in the last 14 days. It will be closed in the next 14 days if no further activity occurs. Thank you for your contributions.

[stale]

This issue has been automatically closed because it has not had activity in the last 28 days. If this issue is still valid, please ping a maintainer. Thank you for your contributions.

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.