This repository contains the Python code for creating a self-driving AI based on the NeuroEvolution of Augmenting Topologies algorithm. NEAT is used to evolve Artificial Neural Networks using genetic algorithm techniques.
- Each generation has 20 instances.
- Each instance is associated with a feed-forward neural network which learns how to navigate the track.
- Each instance is given a fitness value according to the how long it remained on the track.
- New generations are created until at least one car is able to finish a lap without leaving the track.
-
pygame
pip install pygame -
neat-python
pip install neat-python==0.92
The car has a total of 3 sensors, positioned at -45, 0 and 45 degrees respectively. Each sensor has a maximum range of 200 pixels. The blue circles are used to detect if the car left the track (by detecting the green color of grass around the track).
A simple feed forward neural network with 3 inputs and 2 outputs is used. There are no hidden layers. The 3 inputs correspond to the sensor values. There were orignally four outputs: Left, Right, Speed up, Slow down. However, after some experimentation it was found that the later two increased the complexity of the network unnecessarily. The final network has only two outputs: Left and Right.
The following parameters can be modified:
-
config.txt
fitness_criterion fitness_threshold reset_on_extinction pop_size activation_default num_hidden num_inputs num_outputs -
main.py
angle rotation velocity radar_angles translation constant in drive() method of Car class