Turning noise into music: an evolutionary algorithm + crowdsourcing

Professor of evolutionary biology Armand Leroy (Armand Leroi) decided to check how evolutionary algorithms are suitable for generating music. His basic thesis is that a pleasant melody is the result of natural selection (evolution) of sounds. The results of the experiment confirmed that random noise can really turn into music without the help of a composer. Scientific work and samples of generated melodies can be studied on the sites DarwinTunes and Evolectronica .

A computer program randomly selects two small samples from the noise. After that, notes and instruments are randomly selected. Then, using the PerlGP genetic programming system , evolutionary selection begins directly: the original sinusoids “mate” and generate “progeny” (4 samples) with random mutations, from which further evolution proceeds. When the number of samples exceeds 100 pieces, audio files are generated (for details on generating and applying effects, see here ), they are laid out on the Internet, where the public selects the best.


')
The most popular samples are then again driven through an evolutionary algorithm and crowdsourcing. Gradually, pleasant harmonics appear in the hodgepodge of disgusting sounds, which receive more votes and leave more offspring. The quality of music improves rather quickly. When the number of generations exceeds a thousand, then small masterpieces are born with completely unexplained mutations. For example, in Leroy’s experiment, after about 3000 generations, percussion instruments suddenly appeared at the beginning of a melody, although this was not provided for by the algorithm.

However, after some time the average melody score stops growing, that is, the quality of the music reaches a certain maximum level (see the second graph above). According to the authors of the scientific work, this is due to the limitations of the program.

Samples of sounds
DarwinTunes technology
Scientific work