Selection of attributes for modelling Bach chorales by a genetic algorithm

Abstract

A genetic algorithm selected combinations of attributes for a machine learning system. The algorithm used 90 Bach chorale melodies to train models and randomly selected sets of 10 chorales for evaluation. Compression of pitch was used as the fitness evaluation criterion. The best models were used to compress a different test set of chorales and their performance compared to human generate models. G.A. models outperformed the human models, improving compression by 10 percent.