What is the recommended method for using the Experiment Grid API with custom algorithms?

[vrn-sn]

Hi all,

My team is implementing algorithms that don't come pre-implemented in OmniSafe, and we'd love to use the Experiment Grid API to compare them to baselines (and to each other). We can add algorithms to the grid configuration (via the algo parameter), but it appears that this takes a list of strings representing algorithms that are already implemented in OmniSafe (e.g. PPO, PPOLag, etc.).

What is the process for registering a novel algorithm so that its string representation can be used in Experiment Grid for comparison? We have tried registering our algorithms using @registry.register like in the base PPO implementation, but we still run into issues since this file seems to hardcode the list of available algorithms.

Of course, we can manually modify the OmniSafe code, but I would assume there is a better mechanism for this. Thanks!

[Gaiejj]

Sorry for the late reply. You can follow the registration of PPO-Lag. Suppose your novel algorithm is ALGO and it is an on-policy algorithm in naive_lagrange. You can register it in the __init__.py in naive_lagrange, on-policy and algorithms folder. Just like what we have done on PPO-Lag.

[vrn-sn]

Thanks!