Automated Fitness Raters for the GP-Music System

TL;DRAbstract

In previous work, the basic GP-Music System was constructed which allowed human users to evolve short melodies using interactive Genetic Programming. For this project, the basic GP-Music System was improved, and automatic rating of melodies was implemented. The automatic rating is accomplished by automated fitness raters, or auto-raters, which are neural networks with shared weights. The auto-raters are trained using back propagation on a training set composed of sequences and their human assigned ratings. This data was generated during runs of the GP-Music System. Two types of auto-rater were created, one which assigns a 1-100 ranking, and another which indicates which of two sequences is better. The `ranking' auto-rater was able get within 7 of the human rating, and generated some pleasant sounding melodies when substituting for a human in GP-Music runs. The `comparative' auto-rater was never able to get more than 60% accuracy in determining which of two sequences w

Chat with Paper

AI Agents for this Paper