Predictor-surprise pattern

An interesting thought I had a little while back is that a lot of aesthetically pleasing things, like music, art or comedy, seem to rely to some degree on making and breaking expectations. Comedy is maybe the most direct example, because a good joke is usually being surprising or confronting in some way. But in music too there is a constant back-and-forth between expectation and reality. You hear a refrain two times, you're all primed to hear the same thing again, but then it changes subtly the third time. For whatever reason, this seems to be very enjoyable.

It'd be interesting to make a foray into generative music or art by explicitly building an audience model that works as a kind of predictor. The predictor would be constantly searching for patterns in the sequence of notes it's seen, trying to accurately predict what comes next. However, since you're running the predictor and also generating the notes, you can do something a bit interesting: change what comes next if the predictor would have predicted it too easily.

The system would then have some tunable 'surprise factor'. The higher it is the less willing you are to let the predictor win and the more you will subvert its expectations. My prediction is that after a while experimenting you would find some particular value that seems to be the sweet spot for making enjoyable music. But I'm prepared to be surprised.