beaupranisaa / NEATRocket

Training a rocket controller using the NEAT algorithm.

NEATRocket

Training a rocket controller using the NEAT algorithm.

Dependencies:

1. pymunk 5.7.0
2. pyglet 1.5.7
3. neat-python 0.92

Usage:

• Manually control the rocket with your arrow keys.

$ python3 manual.py

• Train the network

$ python3 train.py <name of the directory to save the neural networks to (optional)>

#Example
$ python3 train.py networks/

• Automatic rocket control

$ python3 auto.py <neural network files (optional)> 

# Example:
$ python3 auto.py networks/* networks2/Net_0.p

• Remove all networks

$ bash clear.sh

NEAT Setup

• States/Input:
  • x error = current x position - desired x position
  • y error = current y position - desired y position
  • a error = current angular position - desired angular position
  • vx error = current x velocity - desired x velocity
  • vy error = current y velocity - desired y velocity
  • va error = current angular velocity - desired angular velocity
• Output:
  • longitudinal propulsion states: clamped [-1,+1]
  • top lateral propulsion states: clamped [-1,+1]
  • bottom lateral propulsion states: clamped [-1,+1]
• fitness function : Summation of weighted squared errors of the positional states across time

Contributors:

• Rom Parnichkun
• Pranisaa Charnparttaravanit
• Sitiporn Sae Lim