An Application Framework for Building Evolutionary Computer Systems in Music

Abstract

The authors present a musical composition model that creates rhythmic patterns through a system based on genetic algorithms, involving the interaction of several artificial musicians. In this environment, various composer systems and human musicians may interact within a system based on artificial life.

Numerous projects in recent years have used evolutionary computation techniques for creating art. The works presented in this special section of Leonardo reflect a thriving field of research, with a promising diversity, quantity and quality of contributions. Building upon recent work, we can thus now attempt the construction of global music composition systems capable of creating various types of music. We would like for such systems to be able to adapt dynamically to the tastes of the users interacting with them and to be able to integrate themselves into human culture.

In this article, we address a series of issues regarding the purposes of facilitating this integration and interaction and more closely approaching the dream of creating an "artificial musician." We consider three factors essential for facilitating the creation of a unified work and for achieving the aforesaid goals:

• • elaboration of criteria and guidelines that uniquely define the quality of the works created

• • definition of a developmental plan that allows the establishment of common stages with clearly set goals. This plan must be applied to different works in a consistent way

• • definition of a framework in which the various approaches may co-evolve, allowing their cooperation and competitive interaction.

We will first discuss the necessity of deciding on criteria that define an artistic product's quality. During our analysis of existing works, we have come up with several different approaches to this issue. From our point of view, the quality of a work of art can only be defined through interaction with that work by a group of individuals. The concept of social behavior, which can include relationships among any individuals, whether artificial or natural, within the system, is essential to this type of system.

There is an obvious need for a strategy when planning successive stages in the modeling of human cognitive tasks, given their complexity and the difficulty of tackling the creation of a complex system all at once. Our proposed solution would consist of a series of stages, each of them corresponding to a stage in the history of art (in this case, music). Thus, the first stage would correspond to the most primeval music, and would have features similar to this first kind of music. Each stage would show an evolution with regard to the previous one and correspond to a moment in human history.

Tribe

The Tribe example will serve to illustrate this paradigm. We designed this program in 1999 to work within the musical restrictions that are thought to have existed in the first stages of human music. An anthropological database was used to determine appropriate restrictions [1], leading to two conclusions:

• • primitive music was collectively created by the whole of a tribe or a group within it

• • instrumentally, this music was purely rhythmic, using only percussive instruments.

The Tribe system uses (1) a group of structures that allows it to store musical information applied to the model and (2) a series of mechanisms based on genetic operators for the evolution of the artificial musician.

Structure

The Tribe system is composed of a series of Artificial Musicians' Groups (we will refer to these as AMGs). The AMG represents a tribe of musicians that create a rhythmic theme. The various AMGs compete to create music according to the user's tastes. Each AMG comprises a population or group of several artificial musicians whose common work constitutes the AMG's result. This characteristic has a historical precedent in the collective approach of primeval music. Figure 1 shows the representational structure of one of the AMGs.

Each individual's rhythmical information is independently registered upon a two-dimensional binary grid structure. This grid stores each rhythmic pattern position. An active (marked) grid position indicates...