Towards Co-Evolutionary Musical Improvisation

This paper describes a work in progress on coevolving Artificial Neural Networks (ANNs) for music improvisation. Using this neuro-evolutionary approach the ANNs adapt to the changes in the human player’s music as input, while still maintaining some of the structure of the musical piece previously evolved. The system is called P R I M A L I M P R O V and evolves modules that are composed of two ANNs, one controlling pitch and one controlling rhythm. The results of a quantitative study show that, by only introducing simple rules as fitness functions, the system is able to generate more interesting arrangements than ANNs evolved without a specific objective. The emerging and interesting musical patterns that are produced by the evolved ANNs hint at the promising potential of the system. I . I N T R O D U C T I O N Improvements in computer science, evolutionary computation, and music informatics have enabled many creative performance systems producing music. The goal of most such systems is to exhibit musicality, which is determined by a listening audience or a performing musician. Many diverse approaches have been applied to this domain due to the variety of methods in artificial intelligence (reinforcement learning, evolutionary algorithms, statistical modeling, etc), and to different interpretations and genres of music applied to a specific method [1]. Another goal of artistic expression is to create interesting structures that one cannot immediately imagine. In musical history we find plenty of rules and limitations (for example the 12-tone series), which musicians use to create interesting content. In a way, simple limitations help to support the creative process. Interactive evolution, where human aesthetic judgment is included in the fitness function, has often been used to guide evolutionary creative systems. Another approach is to base the fitness function on music theory to emulate the way a human would play/compose. P R I M A L I M P R O V, while still in its infancy, presents a different approach by introducing very simple rules in the fitness function, which do not necessarily need to rely on musical theory. The most interesting feature is how, even though the system generates music according to rules of our own construction, musical structures emerge that are not obvious to human creators. These structures arise from the freedom we give to the Artificial Neural Networks (ANNs) to musically express characteristics of their topology, thanks to the multitude of ways that the fitness functions can be satisfied. Moreover, by avoiding domain knowledge our system can create music that transcends usual harmony, which can lead to unexpected improvisations. Part of the architecture described in this paper is inspired by MaestroGenesis [2], a tool for computer-assisted composition that is based on interactive evolution. The main difference between P R I M A L I M P R O V and MaestroGenesis is that our system does not require human input, but instead co-evolves its multiple voices according to “primitive” fitness functions. The system presented in this paper is based on the idea of evolving improvisational modules, each capable of creating monophonic melodies. The system can be used to create accompaniments to predefined melodic phrases or, more interestingly, provide an adaptive improvisational companion to a human player. While there are other evolutionary realtime improvisers (GenJam [3], Bown [4]), the proposed system presents a novel modular architecture that allows for the creation of an arbitrary music “instrument”. Finally, when considering the real-time application of P R I M A L I M P R O V, we can note how the feedback loop between the player and the system closes: as the musician plays, he/she finds out how the system reacts to the music played; at the same time the system adapts to the music the human is playing, this change will likewise influence the musician, and so on. I I . B A C K G R O U N D