Overview

This is a simple implementation of a Gym experiment using neat-python

Dependencies

python

gym

neat-python

Running

You can run any of the tests included directly

python exp_name

Test Description

test_neat.py

Running this file will test the python-neat package using an XOR problem.

test_gym.py

Running this file will test the Gym package using the Pole-Balancing environment.

test_gym_neat.py

This file will test the NEATGymExperiment implementation using the Pole-Balancing experiment.

mountain_car.py

This will run an experiment using the MountainCar-v0 environment. It will stop as soon as an agent is able to reach the target in any time.

Example

Running mountain_car.py untill convergence will first display a fitness over generations plot like this one:

Then the winning agent will take control of the environment until you terminate:

NEATGymExperiment

Constructor

Parameter	Default	Description
gym_experiment		The id of the gym experiment.
neat_config		The neat-python configuration.
extract_fitness		A callable that extracts fitness from the experiment record.
runs_per_genome	1	Experiments run for every genome.
interpret_action	None	A callable that infers action from network output.
multiplayer	False	Whether this is a multiplayer game.
server_guide	None	The game_server_guide to set.
n_players	1	The number of active players in a multiplayer environment.
verbose	False	Print extra information.
render_champ	False	Render the champion of every generation.
render_all	False	Render all evaluation experiments.
render_delay	0.005	The delay per frame when rendering (In addition to any delay in the gym Env.)
render_max_frames	200	The maximum frames to render in a test.
mode	'default'	['default', 'threaded', 'parallel]
instances	4	Threads or proccesses, depending on mode
network	FeedForwardNetwork	The network the genome represents (neat.nn.*)

To run the experiment:

winner = exp.run()

The winning champion is always saved in a file with the name experiment_id_winner.dat and the best fitness for each generation is saved in experiment_id_flog.dat