Masters Thesis: A Practical Exploration of Deep Reinforcement Learning

As part of my Masters thesis, I chose a practical way to explore how Deep Reinforcement Learning (DRL) works by building a game and training a neural net how to play...and eventually beat me.

To add to the fun I made it so that a human player could race against the AI!

Highlights on training:

• The agent was trained with over 7 million attempts made.

• The agent was rewarded for how far it traveled through the course and how fast.

• The agent has basic obstacle and ground detection using physics ray casts.

• The course is randomly generated and the course it races the player on has obstacles it has never seen in training.

• I created a secret “advantage“ for the gameplay where the character speed increases the longer they are on the ground. The model was able to “discover“ this advantage

This project made me fall in love with the possibilities of using machine learning in games ❤