Ml-agents: Rewind a Model that Unlearned Something

[GuntherFox]

I'm having an issue where, when I leave an agent learning overnight, I return and it has gotten significantly worse and the cumulative reward never returns to an upward trend. What can I be doing wrong and can I "rewind" the model to its peak state?

[mbaske]

@GuntherFox You can set the --keep-checkpoints option to a high enough value, so that older cptk files don't get deleted. After training, create a backup of your models folder, then edit the "checkpoint" file and set model_checkpoint_path to the cptk file associated with the best results. Now, when you rerun learn.py with the --load option, training should resume from that checkpoint.

[GuntherFox]

Awesome, thanks a bunch! This is exactly what I was lookin' for.

[Sheppio]

Thanks @mbaske, you just saved me a few hours of retraining after pyton threw a wobbly. 👍