This project was done for the class CS5100 - Foundations of Artificial Intelligence during my Master's program at Northeastern University. I secured a GPA of 4.0 for this subject.
I colloborated with my class mate Aditya Shanmugham on this project. Since the major part of the class focuses on search and reinforcement learning, we decided to implement reinforcement learning on our project.
Motivation
By empowering Mario with the supernatural power of Reinforcement learning, we can empower his quality of actions and his love for princess peach. By providing the ability to remember his actions and outcome of his earlier life, we empower super Mario to take calculated risks to reach his goal.
Terminology Map
Since Mario lives in a beautiful vast world with many possibilities where he can explore (states) though he can move in only four directions (actions) there are so many possibilities for him to consider (state-action table). Once he obtains the power of learning from his mistakes in his past life (Q-Learning), we will help him to remember better using Supernatural Power v2 (Deep Q-Learning), where we use a neural network to draw an action using various parameters: Current State, Mario’s Enemies, Power-ups, Princess Peach Position (Final State).
Outcome
The outcome of the project is to compare the efficiency of Q learning vs Double Deep Q learning on Super Mario.
- A detailed report of the comparison is provided in this document. Click here
- An illustrative presentation of the entire project is provided in this slides. Click here
Demo
A simulation of Super Mario trained using Double Deep Q Learning