Modeling MIDI Music as Multivariate Time Series

A method of modeling music using multivariate time series models is described. The models are generated using a hybrid neural networks/discrete particle swarm optimization technique. Such models capture the essence of a piece of music, capable of generating new music sequences. The degree of similarity of the new sequence to the original piece can be controlled by adjusting fitness parameters. A collection of models can be used to represent different sections of one piece or different styles of music. A discrete version of particle swarm optimization was used to manipulate various neural network design parameters, including the lag structure of the input and output variables, the number of hidden nodes, the presence/absence of a linear component, and smooth morphing from univariate to multivariate networks, in order to find the best structure corresponding to the highest fitness of step-ahead predictions against actual data.