Winners in sonic experiment prosper and evolve, losers become extinct
By Paul Ridden
July 3, 2012
For the last few years a new breed has been quietly evolving, gaining strength and complexity as it develops. Fear not though, it's not some monstrous new creation that threatens to devour and consume its way to the top of the food chain. It's something called DarwinTunes and the only folks that may be adversely affected by its continued growth are composers, but not just yet.
Researchers from Imperial College London have built a music creation engine where popular sounds are paired and mated with other successful tunes to create new offspring that inherit features from both parents. Audio loops that fail to please become extinct, parents die off and children get the same rating treatment given to their ancestors. Over time, some rather pleasant electronica has formed without so much as one human composer in sight.
Music is thought to evolve as musicians take what's come before and add their own interpretations, colors and styles to produce something modern and fresh, but little is known about what makes one tune popular and one not among a notoriously fickle music-buying public. To examine the role of consumer selection, a research team led by Dr Robert M. MacCallum (a bioinformaticist in the Laboratory of Immunogenomics at Imperial College London) and professor Armand M. Leroi (from Evolutionary Developmental Biology) has developed a Darwinian music engine populated by a selection of computer-generated audio loops.
In the beginning...
Where natural selection favors the strongest and fittest, the success or failure of the inhabitants of DarwinTunes is determined by popularity with listeners. To avoid any potential cultural and emotional significance caused by the use of preset sounds such as piano, guitar, sax and drums, the project began with the generation of simple computer-generated tones.
Two random founders nicknamed Adam and Eve were mated, with offspring containing random elements from both parents to make new complete individuals. The parents were killed off and the cycle repeated for 100 generations without human interference. After this sample pool was created, the team recruited web-based public consumers to rate the loops on a scale of one to five based on appeal. None of the recruits were aware of the ratings given by others, to avoid any social influence.
A population containing up to 100 eight-second loops was streamed in random order and when 20 loops received a rating, the ten most popular loops were paired up, recombined and had four offspring. The ten rejects became immediately extinct and the successful offspring replaced the parents, which were also killed off.
As the process continued, the researchers discovered that the selection process was making the loops much more pleasurable to listen to. For verification, 2,000 loops were randomly selected from a pool of over 50,000 created during the evolution of the first population, and the consumers were again asked to rate them as with the initial experiment.
In a similar way to determining the fitness of evolved strains of bacteria by comparing them to their direct ancestors, the new ratings were used to estimate the mean absolute of musical appeal of the population at any time. It was found that the mean absolute "increased rapidly for the first 500 - 600 generations but then came to equilibrium. Thus, in our system, musical quality evolves, but it seems that it does not do so indefinitely."
Music information retrieval technology was then used to measure certain aspects of successful loop populations. The results seemed to mirror those for musical appeal, with chord clarity and rhythmic complexity increasing rapidly for the first 500 - 600 generations but then began to fluctuate around a long-term mean.
As a control, a further pool of loops was allowed to run riot without any influence from human preference. The team found the chord clarity and rhythmic complexity of the selected populations to be much higher than those of the unselected control populations by the time generation 100 was reached.
This lack of evolutionary development beyond a certain point is believed to be caused by a decrease in genetic transmission fidelity.
"The reproduction mechanism is not as sophisticated as Nature's," Dr MacCallum told us. The process randomly takes bits from both parents until it has the full complement of digital genes and is able to generate the offspring. Since the possibility of fidelity decrease due to environment was ruled out as loops were identical on all computers, it is thought that "as our evolving tunes got more complex, the nice combinations of melodies, basslines, and rhythms were broken up by this type of mixing (good combinations of say tune and bassline are disrupted and passed on to offspring, for example, receiving two basslines or two tunes)."
"The music evolving on the website now uses a like with like mixing and we'll be interested to see if we evolve more pleasing, complex and more human-competitive music."
When the research paper was published last month in the journal Proceedings of the National Academy of Sciences, 6,931 consumers had registered 85,533 ratings over the course of 2,513 generations of evolution, resulting in 50,480 loops being born.
At the time of writing, the project stands at over 5,000 generations and is still open for folks to take part. I heartily recommend listening to the first basic sine audio medley at zero generations and then the very last - I guarantee you'll be astounded by how far the loops have evolved.
The DarwinTunes team has also just launched a new evolving music stream for 2012 consisting of drum and percussion sounds. Who knows, perhaps some time in the not-too-distant future, offspring from the two experiments will be allowed to mate and produce something that just might be the perfect pop song ...