Abstract
In this paper, we propose the music generation system equipped with a model of human cognitive biases. A musical piece is a construct formed by satisfaction of and deviance to the listener' s expectation. In order to realize the dynamics of expectation, we adopt the loosely symmetric (LS) model that is a kind of distorted conditional probability model based on illogical cognitive biases including the symmetry and the mutual exclusivity bias, proper to thinking and reasoning of human beings. Our system regards features of melodies as the simplest sequence of transitional events from preceding notes to succeeding notes, modifies and generalizes a given melody like human through LS model, and creates new melodies. The average information of the probability distribution to generate new melodies showed that the LS model creates a proper amount of complexity. We conducted an evaluation experiment with more than one hundred participants and found that the melodies generated by LS have a good balance between fulfillment (musical unity) and deviance (unpredictable quality) to expectation. The result indicates the possibility to generate expectation and more human-like melodies by application of human cognitive biases.
