search

TheButlah / makrl

makrl - modular algorithm kit for reinforcement learning
4 stars 1 forks source link

Implement an object framework to enable modular composition of algorithms #5

Closed TheButlah closed 5 years ago

TheButlah commented 5 years ago

To allow us to swap out the underlying model (Tabular, neural net, CNN, etc) and keep it independent of the actual algorithm (PPO, A3C), we need to make an Abstract base class that the different model implementations can extend from. This base class will contain all the methods that the different algorithms can use to interact with the model.

TheButlah commented 5 years ago

I've pushed the basic Agent and Environment objects, but they aren't finished yet.

TheButlah commented 5 years ago

I think I've finished the abstract base classes for Environment, Model, ActionValue, and StateValue in this commit: 7d8d98314a56c5f03d0922647533ce6e05bd5c71. I'm going to figure out how to piece it all together in the abstract base class for Agent. LMK what y'all think.

TheButlah commented 5 years ago

I've made a new issue #10 for this instead, its more specific