Learn To Race

Neural network learns how to drive a car on a track using Evolutionary Algorithms and Reinforcement Learning Simple 2D simulation with pyglet & numpy.

This fork was utilized to write the article Learn to Race: solving with Evolutionary Algorithm and Reinforcement Learning

▶️️ HOW TO RUN?

Install packages

pip install -r requirements.txt

Run main file

Should work with Python 3.0 and higher.

For example:

py -3.10 .\__main__.py

Or use a virtual environment.

Config (Optional)

evolutionary_settings.json

{
	"width": 1280,
	"height": 720,
	"friction": 0.1,
	"render_timestep": 0.025, // time between frames in seconds - 0.025s = 40 FPS
	"timeout_seconds": 30, // maximum time for each gen
	"population": 40, // number of cars
	"mutation_rate": 0.6 // mutation rate after gen
}

reinforcement_settings.json

{
	"width": 1280,
	"height": 720,
	"friction": 0.1,
	"render_timestep": 0.025,
	"timeout_seconds": 30,
	"population": 1,
	"mutation_rate": 0.0,
	"gamma": 0.99,
	"batch_size": 256
}

default_nn_config.json - Default car config for new saves.

{
	"name": "test",
	"acceleration": 1,
	"friction": 0.95,
	"max_speed": 30,
	"rotation_speed": 4,
	"shape": [6, 4, 3, 2], // neural network shape - do not change first and last layer
	"max_score": 0,
	"gen_count": 0
}

ddpg_config.json

{
	"name": "DDPG",
	"acceleration": 1,
	"friction": 0.95,
	"max_speed": 30,
	"rotation_speed": 4,
	"shape": [6, 32, 32, 2],
	"max_score": 0,
	"gen_count": 0
}

td3_config.json

{
	"name": "TD3",
	"acceleration": 1,
	"friction": 0.95,
	"max_speed": 30,
	"rotation_speed": 4,
	"shape": [6, 32, 32, 2],
	"max_score": 0,
	"gen_count": 0
}

🏎️ Environment and Trach Generation

Evolutionary Algorithm

Write "e"/"evolutionary" when running the app to choose the evolutionary algorithm to train the MLPs

Reinforcement Learning

Write "r"/"reinforcement" when running the app to choose the reinforcement algorithm to train the MLPs

DDPG and TD3

Write down "d"/"ddpg" or "t"/"td3" to choose which algorithm you want to utilize to train.

⚖️ LICENSE

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.