IT Carlow Project Showcase: Computer Games Development Year 4

Acadamic Year: 2020/2021

Student name: Jonelle Lawler

Student number: C00205084


Project title: A comparison between genetic algorithms and backpropagation for training a neural network to play a game.

Quick description

The aim of this project was to train neural networks using both genetic algorithms and backpropagation to play the game Frogger and compare them

in three key areas: effectiveness, efficiency and ease of use

My initial idea was to train the networks to play Backgammon

This would have been interesting as the game is non-determinisitc. How could I measure the effectiveness of the algorithms when luck was integral to the game?

An issue with this unfortunately became clear over time. I would have the play the game possibly thousands of times in order to train my networks and this is just not possible to do over the course of a few months

Research indicated that genetic algorithms score higher in all three areas

While backpropagation can work very well with well for some projects, it can struggle with unexpected data

My personal experience supports the hypothesis that genetic algorithms are superior when it comes to ease of use only

I developed this project using the Unity engine with C#

Click here to read my project report

Click here to view my Github

Click here to watch a video presentation of my project