Target-driven Genetic Algorithms for Synthesizer Control

A set of experiments is described which use Genetic Algorithms (GAs) to find the synthesizer parameters required to synthesize a target sound. Comparisons between the target and candidate sounds (for the algorithms’ fitness functions) are calculated via a mapping from the sound data to a set of timbral, perceptual, and statistical sound attributes, many of which have been used in recent machine learning research. In particular, the performance of a new type of modification to the standard GA, using an Increasingly Discriminating Fitness Function, is evaluated; the performance of a weighted-attribute fitness function is also tested. The main results are that all GA techniques perform better than random search, but that the new variations fail to provide an improvement over the standard GA.