N-gram fitness function with a constraint in a musical evolutionary system

This paper describes an evolutionary music composition system that combines trainable music critics with a bag of notes constraint. Unlike many evolutionary composition systems, there is no human interaction involved in the loop. The role of the human within our system is to select the set of melodies to train the critics on, and choose the bag of notes. The trainable critics are N-gram model. The system then evolves pleasant sounding melodies by permuting the order of the notes selected from the bag. The bag of notes constraint prevents a previously observed problem. We solve a prior problem where the maximum likelihood sequence (MLS) generated by our N-gram model is repetitive using the bag of notes constraint. In this paper, two experiments are constructed. Both experiments are identical except for the melody representations, which are absolute pitch (AP) and pitch difference (PD) representations. In each case, the bag of notes constraint results in more pleasing melodies, and the melody shape of the melodies are more controllable and predictable with appropriate operators. We suggest that using absolute pitch representation leads to better sounding melodies than the pitch difference representation, especially if the user enjoys arpeggio effects.