Neural network fitness functions for a musical IGA

This paper describes recent enhancements to GenJam, a genetic algorithm-based model of a novice jazz musician learning to improvise. After presenting an overview and update of the current interactive version of GenJam, we focus on efforts to augment its human fitness function with a neural network, in an attempt to ease the fitness bottleneck inherent in musical IGAs. Specifically, a cascade correlation technique was used with data taken from populations of musical ideas trained by human mentors interactively. We conclude with a discussion of why this approach failed, and we speculate on approaches that might work.