Algorithmic composition, artificial intelligence

[ilSignorCarlo]

Hi,
I am going to work on a project for an Artificial Intellingence course.

I want to create a software to compose music starting from some other music (possibly in a MIDI format) given in input. My idea is to use the input to reproduce music in a similar style.

I think I can use Markov Chains for this particular purpose.
My only requisite is that I have to use some form of knowledge representation (this is what the teacher asked me), but I'm not really sure what to represent, and how.

During the course we studied a bit of Lisp, so I thought to use Common Lisp Music (for example), but this is not really important, I think.

Do you have any suggestion about some kind of knowledge representation in the music programming field?

Is it possible to use ChucK with Lisp, for example?

[DrJustice]

ilSignorCarlo, perhaps you should head over to the ChucK programming language forum. I'm sure someone there will know about any ChucK and Lisp connections.

DJ
--

[CJ Miller]

ChucK is really interesting, but I am still too busy and a little intimidated to get into it now. By far the easiest algorithmic manipulation of MIDI data I have done was with IRCAMs OpenMusic. It is written in Common LISP and interfaced with Midishare. Unlike ChucK it is not much of a real-time performance environment. For composing and trying different iterations it's a fun and powerful thing.

latest code release:
http://sourceforge.net/projects/ircam-openmusic/

main page, including documentation:
http://recherche.ircam.fr/equipes/repmus/OpenMusic/

[Inventor]

I have experimented a little bit with neural nets and ChucK. I make the nets do processing of the music in real-time and that is computationally intensive so I've been limited to about 50 neurons. Still, that is enough to do some song learning in a very crude way.

In one example I gave each neuron a variable delay element and used a stochastic optimization algorithm for the training. In another i trained with backprop on the error of the song's FFT coefficients. Both techniques start out with the net just clicking or babbling, then gradually real sound starts to come out. Once the training is done, I degrade the input signal and see how well the output holds up.

I'm not sure how much of these programs is based on learning the song and how much is just learning to pass audio thru to the output. You'll want to set up your simulation so that simply passing the audio from input to output is not a solution to the learning problem.

Regarding your question, just what do you mean by knowledge representation? I'm not familiar with that phrase.

I don't know much about markov chains, but from what little I do know, they would be just great for this purpose. Let us know what you end up creating, I like AI music programs!

[ilSignorCarlo]