Which Reinforcement Learning Framework is the Best?

preview_player
Показать описание

We compare 10 most popular Deep Reinforcement Learning frameworks in this video.

We scored each frameworks on the following metrics.

1. How actively is the framework being developed?
2. Does it have the most important features?
3. How modular is the framework?
4. How good is its documentation?
5. Is the community or ecosystem strong?

Our conclusion is that Ray Rllib is currently the best Deep RL framework, closely followed by Acme, TF-Agents and Tianshou.

The videos in the course build on top each each other. For optimal results, I recommend watching the videos in order, starting from the beginning of the playlist.
Рекомендации по теме
Комментарии
Автор

SB3 is IMO the clear better choice for most small/medium scale projects. Excellent docs, easy to ask for help for, and fast. Ray really only makes sense for large scale training on clusters. And their docs feel a lot more cluttered and messy in comparison

GeneralKenobi
Автор

Thank you very much for this video. As mentioned in the video, could you please add a link in the video notes to show the details of the scoring system?

davidschnermann
Автор

Unfortunately, this comparison falls apart when considering different tasks. For example, Ray/rllib is a terrible choice for control of real dynamical systems (robot etc.). It's really been over engineered for the toy environments from gym.

I think a lot of people forget that reinforcement learning as an academic game is completely separate from what has been working in practice. I would say sb3 or rsl_rl are better starting points. But it will always depend on your task.

georgehart
Автор

Hey dibya, where does openai’s gym fall in this category of frameworks? Can it even be defined as a framework or not?

joelvasanth