Cartesian Genetic Programming Parameterization in the Context of Audio Synthesis

This letter presents an evaluation of the effects of elitism, recurrence probability, and prior knowledge on the fitness achieved by Cartesian Genetic Programming (CGP) in the context of DSP audio synthesis. Prior knowledge was introduced using a probabilistic learning method where the distribution of nodes in the expected solutions was used to generate and mutate new individuals. Best results were obtained with traditional elitist selection, no recurrence, and when prior knowledge was used for node initialization and mutation. These results suggest that the apparent benefits of recurrence in CGP are context-dependent, and that selecting nodes from a uniform distribution is not always optimal.