An agent capable of playing various games in Gym environments, using NARS for planning
This repository hosts both a library that facilitates interfacing OpenNARS for Applications with OpenAI Gym and a bunch of experiments using Griddly.
You need to have OpenNARS-For-Applications installed somewhere. It is expected that you will export a NARS_HOME environment
variable to specify where it is built.
The recommended way to install dependencies is with Poetry: poetry install. poetry shell will open a shell with the virtual environment activated.
Additionally, dependencies are listed in the env.yml file, that can be used to create a Conda environment (might be outdated).
lock.yml is a lock-file that you can use to make a reproducible environment, Conda might fail to create it though.
Griddly requires to install Vulkan SDK, refer to its README.
Install in editable mode to your environment: pip install -e .
There is also a PyPI package available: pip install narca.
There are multiple experiments under the experiments directory. Each one can be launched with python <path to the script>. You can observe the rendered environment
and run TensorBoard to see performance metrics. TensorBoard run data gets saved to runs.
There is also Neptune.ai integration that you can use, you need to set NEPTUNE_PROJECT and NEPTUNE_API_TOKEN.
Each Griddly experiment has its own subdirectory, each has at least the main experiment - this is supposed to hold the currently established simplest best performing agent.
Besides the usual main experiments there are some interesting variations as follows:
experiments/zelda/procgen: uses a level generator that places a key and a door randomly around the level.experiments/drunk_dwarf/curriculum_learning: tries to find out whether there is a measurable gain in learning efficiency with a curriculum learning setup.
Here you can watch the presentation of some results from this project as of August 2022: https://youtu.be/ciDH3oJbP0c?t=20868