Alternative to online training

[eliofreitas]

Greetings, First let me give you all a praise for this amazing library. Now lets follow unto my problem. Recently I have been developing my thesis around behavioural cloning using this package. I was going to update my branch to the latest but I stumble upon that the online_training mode was deprecated. I found an issue (https://github.com/Unity-Technologies/ml-agents/issues/2901) talking about it saying it was limited, because it could only learn from example and that's the justification for deprecating, I understand the need for deprecation however the complete elimination sounds like a hard decision without having an alternative at hand at least for the time being.

As for my thesis for what I was checking I can rollback to 0.10.0 for now. However, for future developments will there exist any alternative?.

In the other issue (https://github.com/Unity-Technologies/ml-agents/issues/2901) you talk about using Gail is that something that can be done or will be done?

[chriselion]

Hi @eliofreitas, Unfortunately we don't have a timeline for if/when online imitation learning will be brought back. I added your feedback to our internal tracker (MLA-422).

[eliofreitas]

Understood. Thank you for the response.

[Ro-tico]

@eliofreitas have you worked around this problem?

[github-actions[bot]]

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.