Snake AI 🐍

An evolutionary algorithm that ajusts the weights of a neural network to control a snake.

Prerequisites

• Kivy

  • See instalation instrutions https://kivy.org

• Install pip dependencies

  • pip install -r requirements.txt

Running the program

python src/main.py

• In the file main.py at lines 123, 124 and 125 we have the variables USE_GUI, USE_SAVE and MANUAL.
  • USE_GUI enable the visual interface.
  • USE_SAVE makes the program use the file save.pkl as the Snake AI.
  • MANUAL let you play the game! 😄

AI Learning

• 1° Start #population_size individuals of neural network with random weights.

• 2° The population of neural networks are evaluated and ranked by their fitness.

  • The fitness of each neural network is calculated by analysing the movements of the snake.
    • Eat an apple, fitness += 40
    • Move farther the apple, fitness -= 3
    • Move closer the appple, fitness += 1
  • It's important to give a high value for eating the apple because sometimes the snake needs to make a distant path to reach the apple and that will compensate the negatives values given by moving father.
  • Moving father needs to be higher in module than moving close to make the snake stop running in circle.

• 3° Remove some percentage of the population worst individuals and add new ones.

  • It's important to recreate some indivuals because sometimes to reach a higher fitness it's needed to make a bigger change in the weights which it's not achievable using crossover and mutations (maximum local).

• 4° Make the crossover.

• 5° Make the mutation.

• 6° Start a new generation.

• Observation: One of the evalution problems is the noise given by the location of the apples. Sometimes the apple can be in better place and this may give a higher fitness incorrectly. To reduce this noise and improve the fitness acurracy, it's important to evaluate the same network more than one time (play more games).

Team

Paulo Pinheiro Lemgruber Jeunon Sousa
Matheus Carvalho Nali