Generation of Musical Sequences with Genetic Techniques

Biologically inspired computational methods have recently attracted much research interest in the field of computer music. This group of methods is a subset of computational intelligence, and is represented by neural networks (NNs), genetic algorithms (GAs), and genetic programming (GP). A common feature of these methods is that they all mimic biological processes: NNs realize the evolutionary device of learning from experience, as humans and other animals do, while GAs and GP are based upon procedures that imitate the laws of natural selection. Genetic algorithms have been shown to display distinct performance improvements compared to enumerative, calculus-based, and random searches of a given arbitrary search space (Goldberg 1989). This is achieved by combining aspects of these search methods to result in a " guided random " search. The genetic algorithm samples points throughout the search space for their worth, and is " blind " to any information regarding the search space apart from this measure of worth. This makes genetic search techniques more general and applicable to many search or optimization tasks, as long as an appropriate encoding scheme is employed. By using a population of candidate solutions, rather than single individuals, an inherent parallel-ism in the search process is apparent. This is because the search for an optimum solution is " performed over genetic structures (building blocks) that can represent a number of possible solutions " (Filho, Treleaven, and Alippi 1994). Genetic programming techniques expand on the versatility of GA techniques by evolving generations of functions, rather than representations of a single function. This is based upon the need to evolve computer programs to solve a problem, rather than evolving solutions to a given fixed problem (Koza 1992). Computational intelligence techniques have been applied to musical problems across a wide range of subject areas, including algorithmic composition , artificial listening, musical cognition, and sound synthesis (Todd and Loy 1991; Balaban, Ebcioglu, and Laske 1992), while GAs have been applied principally to the musical tasks of composition and synthesis. The search space for both of these tasks is potentially vast, considering the number of possible musical compositions and musical sounds. The use of stylistic or timbral constraints upon a composition or synthesis task reduces the size of the search space, yet the space still remains inefficiently large for traditional search techniques. Genetic techniques are able to perform well with large search spaces, owing to their inherent parallelism. The aim of …