Any possibilities accessing values from the agent's brain?

Hi!

Is it possible to access values from the barracuda network itself?
I'd for instance like to find out what the action output probabilities are before they're sampled during inference.
Or maybe even access the entire last layer?
Any ways to do this?

Thanks!

 

Hi @MarkTension,
The short answer is we don't currently have an interface in ML-Agents to support this. You could theoretically do this with Barracuda, but it would be outside of the model execution we do within ML-Agents.

Could you create a formal request on our GitHub repo and describe exactly what you want there. This way we can track it and give you updates accordingly.
Cheers,
Chris

 

Think you're posting in the wrong thread

 

Yup, sorry about that.