CST-RL is a framework for building assynchronous RL agents as part of a larger cognitive architecture through the Cognitive Systems Toolkit (CST). Its focus is largely on solving the structural and communication challenges of an assynchrounous RL agent rather than the specifics of the algorithms implemented. As such, while a couple of benchmark algorithms are implemented directly in Java, the plan is to implement more state-of-the-art algorithms through the use of external libraries.
Moreover, this repository uses the open-source game engine Godot to build complex environments through the use of the CST-Godot framework.
This repository is divided in the following sections
Includes a simple Python script that uses NumPy and Pyplot to generate graphs from CSV files. This is used to generate graphs of cummulative reward over time of different algorithms from the CSV files generated by CST-RL.
The Godot project used as environment for the RL algorithm. It communicates with CST through the use of HTTP requests.
The framework itself. Here, it runs as a web API that communicates with a Godot environment, receiving percepts and returning actions.