Algorithms

CREW takes every effort to make the algorithms independent of the environments provided, giving as much flexibility as possible to researchers for defining their own algorithms and using whatever libraries they are familiar with. ML-Agents, which centers around this functionaility, wraps up the Unity environments and provides standard interfaces that are common in reinforcement learning research. Additionally, the feature of information channels allows for more diverse feedbacks that researchers may feed to algorithms, leaving space for types beyond just binary rewards.

In this part of the documentation, we introduce how to use the interfaces to communicate with the environments from CREW.

We then provide human-guided examples implemented by torchrl. In principle, there is no restriction for what libraries to use nor how an algorithm is implemented. It's aimed that the examples get users familiar with the functionalities of CREW and help users to implement their own algorithms with ease.