Measuring custom metrics

How severe does this issue affect your experience of using Ray?

  • High: It blocks me to complete my task.


I’m using RLlib to train my agents on an environment. I want to collect some metrics about their behavior on every training step. I notice that when I run ppo.evaluate(), I get a dict with all kinds of useful metrics (like the reward for each agent) and there’s also a dict called custom_metrics which is empty.

I hope it won’t be too presumptious to assums that the custom_metrics dict is intended for users to add their own custom metrics. My question is: How do I do that? How do I give RLlib a list of my custom metrics and the logic for measuring each one of them?

Thanks for your help,
Ram Rachum.

Sorry, I just found the relevant documentation page: Training APIs — Ray 2.0.0

Please disregard this thread.


1 Like