lenghuang/atla-agt

A visualization of a Regret Minimization Learning algorithm for Two Person games, but Avatar themed! 15-251 Fall 2020 Project

Regret Minimization in Avatar

15-251: Great Theoretical Ideas of Computer Science

Hey there! My name is Len, and this is my submission for a class I was taking in Fall 2020 commonly called "251". We got to choose a topic and do really whatever, so I chose to explore the topic of algorithmic game theory. More specifically Regret Minimization. Combined with the iconic rerelease of Avatar the Last Airbender onto Netflix in 2020, I found inspiration to combine the two in this project.

Check out the live site here: https://atla-agt.web.app/home

Directory

• code: Contains various implementations of regret matching, as well as a flask API that is being hosted with pythonanywhere.com.
• notes: Contains conceptual information and related notes.
• pictures: Some of my drawings to better discuss AGT
• site: React website to put everything together

Visualize the Algorithm Over 50,000 Iterations

Check out how the strategies of the players change after 50,000 iterations of the regret matching learning algorithm.

Test This Out On Your Own "Games"

Input your own tables, or two-person normal form games, with their respective utilites to try out the learning algorithm on your own games.

Sources

• An Introduction to Counterfactual Regret Minimization
• Faster Regret Matching
• MIT 15.S50 Game Theory and Analytics
• Adaptive Procedure Leading to Correlated Equilibrium