Dynamic Techniques for Genetic Algorithm–Based Music Systems

Genetic algorithms (GAs) are widely used in the worlds of computer music and art. These algorithms are well suited to problems in which a search for an optimal candidate encounters maxima within a complex landscape defined by aesthetics. In such cases, which often are ill conditioned, optimization is not always of primary importance, as software developers and musicians appear to be more interested in exploration. This article highlights the primary deficiencies within this special relationship between creativity problems and GAs. This is achieved through the analysis of results obtained from an extended series of experiments undertaken on the fitness function. A dynamic genetic-algorithm variation that has been adapted to solve creativity problems can then be recommended. Programmers must define their goals and let the system dynamically allocate the main parameters of the GA to serve these goals.

The core of this new algorithmic design is the implementation of an innovative idea to simulate a damped oscillator to calculate the fitness-function coefficients. The integration of the damped oscillator into a simple genetic algorithm potentially constitutes a challenge: the construction of an environment where composer–programmers will be able to structure their algorithms based on a set of targets and a valid data representation, leaving the fitness function to be decided dynamically by the system. The accomplishment of such a goal initiated the motivation of the work presented in this article.

Genetic Algorithms and Creativity Problems: A Special Relationship

GAs are based on the computer simulation of the evolutionary ideas of natural selection and genetics. According to this concept, a pool of candidate solutions, known as individuals or phenotypes, corresponds to a population of abstract representations, commonly called chromosomes. The population is stochastically generated, and it evolves for a number of iterations called generations. In each generation, the fitness of every individual in the population is evaluated and assigned with a fitness value that resembles the distance of each individual from the desired target (a maximum or a minimum, for instance). Then, based on their fitness, multiple individuals are stochastically selected from the current population (selection), recombined (recombination or crossover), and randomly modified (mutation) to form a new population. The new population passes through the same stages of selection–mutation–recombination until a maximum number of generations has been reached, or an individual has produced a satisfactory solution.

Holland presented the genetic algorithm in 1975. After the algorithm was tested and discussed explicitly (Goldberg 1989, 2002; Davis 1991; De Jong 2001), it was soon applied in a plethora of problem domains. There was a wave of applications developed in the music-creativity domain, especially after the mid 1980s, when the computational capacity of the desktop computer was significantly increased. Variations (Jacob 1994), MusicBlox (Gartland-Jones and Copley 2003), CONGA (Tokui 2000), and GenJam (Biles 1994) represent a small sample of such systems, and the current list of GA music systems is increasing rapidly. A more analytical investigation of these applications is presented in Assayag (1998) and Burton and Vladimirova (1999). The establishment of the relationship between GAs and music-creativity problems can be explained on the basis of two characteristics of the behavior of GAs: rapid search-space exploration and multiple optimization, as discussed in the following section.

Search Space

The search space in music-creativity problems is vast. This excludes most common search algorithms, as it would take an infinite time to explore. In addition, the nature of the landscape is unique: While climbing a hill of the landscape, a slight perturbation can throw the exploration to a minimum (Jacob 1994), completely disorienting the algorithm. Traditional optimization mechanisms fail to locate a sufficient solution as the landscape is viewed in a tree form, where the algorithm searches down to the end of a pathway before trying an alternative starting point; a breadth-first approach follows until all available options are tested or a satisfying solution is reached. Added heuristics may improve the efficiency by limiting the search space, but at the same time, they restrict the exploration in significantly smaller search spaces (Gartland-Jones and Copley 2003). GAs appear to have the skill...