This repository is a fork of https://github.com/simondlevy/neat-gym containing code allowing you to train, test, and visualize Gymnasium environments (games) using the NEAT algorithm and its variants.
The two goals of this project are
-
Make this work as simple as possible, via config files.
-
Make the code run fast, by simultaneously evaluating the fitnesses of the population on multiprocessor machines.
To get started you should install neat-python and PUREPLES from source. Then do the following:
% python3 neat-evolve.py config/cartpole
This will run neat-python on the CartPole-v1 environment using the parallel fitness evaluator, so you can take advantage of all the cores on your computer.
Once evolution finishes, you can try out your evolved network by doing:
% python3 neat-test.py models/CartPole-v1<fitness>.dat
where <fitness> is the fitness of your evolved network.
The visuals folder will contain a PDF showing the corresponding model,
and the runs folder will contain a CSV file with the history of the
fitnesses (mean, standard deviation, max). To visualize this history you
can run the neat-plot.py script on this CSV file.
To train spaceinvaders-RGB (https://gymnasium.farama.org/environments/atari/space_invaders/), run
% python3 neat-evolve.py models/spaceinvadersRGB
To train spaceinvaders-RAM, run
% python3 neat-evolve.py models/spaceinvadersRAM
Info on how to set up the config files for atari games can be found in the README located in the config folder
IMPORTANT
Running gymnasium games is currently untested with HyperNEAT and ES-HyperNEAT, and may not work.
NEAT-Gym supports HyperNEAT via the --hyper option and
and ES-HyperNEAT via the
--eshyper option.
There are two ways to specify the substrate:
-
In the
[Substrate]section of the config file (default) -
Via a
get_substrate()method in your environment. This method should return a tuple containing the input, hidden, and output coordinates and the name of the activation function (e.g.,sigmoid). For ES-HyperNEAT, the hidden coordinates should beNone.
IMPORTANT
Running gymnasium games is currently untested with Novelty Search, and may not work.
NEAT-Gym supports
Novelty Search
via the --novelty option. To use this option, the info dictionary
returned by your environment's step() method should have an entry for behavior,
whose value is the behavior of the agent at the end of the episode (for
example, its final position in the maze), or None before the end of the
episode. For an example, try out the
maze
environment.
The neat_gym module exports two functions:
-
read_file loads a .dat file from command-line arguments, returning the evolved network and the name of the environment used to evolve it (as well as flags for recording a move and suppressing the display).
-
eval_net takes the output of
read-fileand runs an episode of the environment using the network.
To make these functions available other projects (e.g., for 3D animation of your environment during testing), do:
% sudo python3 setup.py install
You will also need to create a NEAT configuration file for your environment. As usual, the easiest way to do this is to take something that works (like the config file for CartPole-v1) and modify it to do what you want.