A quick exercise to recall Genetic Algorithms. Teaching a three-polygon figure to jump over the ball. Recreation of this video without using a neural network. The genetic algorithm tries to find a sequence of moves needed to minimize the fitness function which is the overlap between the obstacle and the figure. Physics engine: pymunk, display: pyglet.

figure.py - A geometric shape out of three polygons that moves its left/right polygon up/down

universe.py - Static space with 4 walls, obstacle, and a figure

simulation.py - Simulates actions of a figure and a moving obsticle (30 fps and 4 actions per second)

display.py - Recreates a simulation and displays it in pyglet

ga.py - Genetic algorithm. Generates random lists of actions, performs uniform crossover, evaluates and picks elite individuals.

• Individual is represented by its actions (0: no movement, 1 and 2: left polygon up and down, 3 and 4: right polygon up and down). For example: [1, 0, 4, 3, 2, 0, ...] - left poly up, nothing, right poly down, ...

• Crossover each bit is chosen from either parent with a given probability (uniform crossover)

• Fitness function is a sum of overlap and a penalty of making a move

• Simulation takes 8.6 seconds (untill the obsticle hits the left wall) and the figure makes a move every 0.25 seconds

pip install requirements.txt
cd genetic_algorithm
python3 main.py