How many times you were playing a good game and after spending a few hours you were greeted by a screen saying game over? And after spending a decent amount of money there is nothing what you can do with your Blu-ray disc, right? We have a solution to this problem, a recurrent neural network that is able to generate new levels for your game!

What it does

Based on experience, a set of Mario Bros maps, the network generates new levels. The only thing we require from you is a length of a map as well as the crazy factor. Cause why keep stuff boring?

How we built it

We built this using Keras, a python library for deep learning networks. We first took maps available online, turned them into ASCII representation using OpenCV and then feed the text as the training set for the network. After generating a new ASCII map we turned them again to Javascript format, which is understandable by the version of Mario we are using.

Challenges we ran into

At the beginning we were trying to build a network able playing Mario itself, unfortunately we had to leave this one, since it very hard to setup the feedback step. We focused on the generating new maps only.

Accomplishments that we're proud of

It is working!

What we learned

A better understanding of recurrent neural networks. <3

Built With

Share this project: