search

IntelLabs / coach

Reinforcement Learning Coach by Intel AI Lab enables easy experimentation with state of the art Reinforcement Learning algorithms
https://intellabs.github.io/coach/
Apache License 2.0
2.32k stars 460 forks source link

Understanding tensorboard outputs of rl-coach model #381

Open Eriz11 opened 4 years ago

Eriz11 commented 4 years ago

Hi all,

Related the issue that was closed some days ago ( #374 ), the layers output is somewhat confusing when printing the graph. This also extrapolates to the outputs seen on tensorboard when saving the logs via VisualizationParameters param of tensorboard.

I think that having some guidance in the docs or in the naming of the different layers/components will be crucial to understand and interprete if the agents are learning correctly and if there could be some tweaking that could help them learn better.

¿Any insights about this @galnov @gal-leibovich and etc?

As always, I'm keen on helping as much as I can once I know more (rewriting docs or whatever).

Thanks in advance,

Eriz11 commented 4 years ago

¿Anyone that could make some input to this?

I would like to be able to understand if my model is converging, if the training process is correct and etc, which seems to be very tied to tensorboard outputs; however, in complex models like Rainbow, I don't see a clear path on how to assess this.

Please, if you have any idea of how to do this or how you are doing it actually with the implication ofrl-coach, it would be very helpful for me.

Thanks in advance,