Genetic Algorithm Plays Pong

By Elliot Kirk and Colin Dutter

Genetic Algorithm in action

Prerequisites

• Python3
• Pygame

How to Run the Neural Net Breeder

• From the directory you cloned this repo into, run python3 breeder.py
• There are a couple of optional arguments you can include:
  • -g, int: specifies the number of generations to evolve over
    • default = 5
  • -p, int: specifies the size of each population (must be even)
    • default = 10
  • -load, string: specifies a specific genome (located in the /genomes folder) to start a new generation from
    • default = None
      • the user can provide a list of names to load multiple genomes at once (ex. -load '[Name1, Name2, ...]')
  • -strict, boolean: specify whether or not you want strict breeding (use a fitness function based only on scoring points, excluding the number of times the individual hits the ball)
    • default = True for the first generation, False for subsequent generations
  • -arena, boolean: enable the dreaded arena mode; a free-for-all four-paddle version of pong
    • default = False

How the Genetic Algorithm Works

This program uses a neural net in combination with a genetic algorithm to find the ultimate pong-playing AI.

The genetic algorithm initially produces a random list of weights (floats in the range [0, 1]) and uses those values as the weights on the synapses of a neural network.

At the end of each generation, the fitness of all the members of the population are evaluated

• if no individuals score a point, a new population is generated, strict_breeding is set to False, and the generation count is set to 0,
• if only one individual scores a point, random mates are created and crossed over with the winner to create a new population
• if more than one individual scores a point, those individuals will mate and create a new population from their genomes

If the population makes it past the second generation on their own, they start to play themselves rather than the basic AI
Once the generation count hits the value specified in -g, the simulation ends and the fittest individual's genes are saved into /genomes

In-Game Controls

• Up moves the player controlled paddle up
• Down moves the player controlled paddle down
• Left slows down the game
• Right speeds up the game
• R resets the current game
• P pauses the current game
• Space allows the user to take control of the paddle on the left

Resources

If you are interested in this project but don't know where to get started, don't worry, we were there too. To get started we did some research and came across a few good resources online.

• Pong Neural Network
  • This tutorial teaches you how to write a pong-playing AI using neural network libraries, rather than from scratch
• Neural Network plays Pong
  • Very similar project to this one, but written in Java