The time for Project AI is swiftly approaching and I would like to do a custom project instead of the pre-defined ones. To make that happen, two things are needed:
- A fellow student willing to team up with me to work on this project.
- The approval of one professor who will supervise the effort.
What is this project about? It's a procedurally generated music game. So, you give it a music track as input and it generates a game level that is in sync with the music. While playing, your interaction with the game is in accordance to the music. The more you follow the music, the better.
More information after the jump.
How will it work? The basic idea is to split the work into two parts: The first part is extracting, using machine learning, a sequence of button presses from the music. Analysing the audio to get the best features is going to be important in this part. The second part is building a game level around those button presses. In essence, you have to place level geometry and entities so that to give an incentive for the player to follow the button presses (which in turn means to follow the music). This means that the gameplay must be designed in such a way as to promote this. Then you define a search problem to find a good level that conforms to the restrictions imposed by the gameplay and the button presses.
What is done so far? This thing started as a course project but ultimately spiralled out of control as I continued building on it in my free time. What we have so far is a "proof of concept", which includes extracting button presses in a very naive way (frequency analysis for features plus a linear SVM plus a dataset of 10 songs), a few iterations on the gameplay (one-button platformer with gradually decreasing infinite double jumping) and a simple solution to the search problem (a level representation based on placing at most one gameplay element at each position in the x-axis, with the position of the element on the y axis being pre-defined; this is fed to a genetic algorithm that tries to find the best level based on some criteria). Finally, there's a draft paper that describes how this whole thing works so far. The original intention was to publish this as a "proof of concept" and then allow for subsequent papers to describe further improvements.
What's next? Nothing is set, everything is possible (okay, maybe not everything). Does this project sound interesting to you? Leave a comment (or drop me an email) and I'll send you what I have so far. Then, you can decide if you'd like to work on this. If you do, we'll try to get a professor on board and after that we can decide the specifics of what will be done.
Count me in...
ReplyDelete