π Emoteο
Emote β Embarkβs Modular Training Engine β is a flexible framework for reinforcement learning written at Embark.
Installationο
Install PDM following the instructions on the PDM site. Then install the package using
pdm install
Currently, our supported version of PDM is 2.10 β other versions may work, but there are no guarantees.
Ideas and Philosophyο
We wanted a reinforcement learning framework that was modular both in the sense that we could easily swap the algorithm we used and how data was collected but also in the sense that the different parts of various algorithms could be reused to build other algorithms.
emote is a torch implementation of embarkrl
Emoteο
In order to do reinforcement learning we need to have two things: A learning protocol that specifies which losses to use, which network architectures, which optimizers, and so forth. We also need some kind of data collector that interacts with the world and stores the experiences from that in a way which makes them accessible to the learning protocol.
In Emote, data collection is done by Collectors, the protocol for the learning algorithm is built up of Callbacks, and they are tied together by a Trainer.