How to evaluate actors trained in custom environments in rllib (Ray)?

Question

Following the SimpleCorridor example I can create my own env and train a model! nice. But when I try to evaluate this trained model, rllib does not recognize my custom env.

How can I evaluate a trained model on a custom environment?

When I use rllib rollout ... like its suggested here it does not recognize my env because it's custom one. I was hoping to have a function like run_experiments but like evaluate_experiment so I can call it on my project inside one of my files.

Thats the issue. If you want to see my custon_env is this one

Right now I'm having to copy my environment and paste it inside the gym/envs/ package directory and register it in the __init__.py file.

Would be good to see another way to do this

Thanks

Answer 1

You can copy the rollout.py script and modify it to register any custom envs or models in the start of the script. That will allow it to work with any envs not on gym already.