Screencap of live play
We wanted to create a game that helped us further understanding some main methods of machine learning.
What it does
This game encompasses a machine learning model that learns how to play a simplified version of the popular game, Super Smash Bros. It uses a population of individuals with neural networks that evolve over time using a genetic algorithm.
How I built it
Each machine learning algorithm is implemented in native java using Eclipse.
Challenges I ran into
Implementing the genetic algorithms to evolve individuals who already have neural networks within them was difficult. We solved this by abstracting the neural networks and treating every individual solely according to its fitness value calculated from the neural network within it.
Accomplishments that I'm proud of
We used no libraries to implement machine learning models from scratch. This is not typically done due to the complex nature of such algorithms--we learned a lot about the inner-workings of genetic algorithms and neural networks by doing so.
What I learned
We gained a deeper understanding of Genetic Algorithms and Neural Networks by cross implementing them. We also learned about the nuisances of having 2 ML algorithms in a single model.
What's next for MITSSAI
We hope to further customize the program with more platforms and maybe even levels.