Sigma Alpha Iota

rhythm & algorithm Compositional Artificial Intelligence By Jessica Keup W hat do you think of when you imagine music written by a computer? Something boring, plain, and uninteresting? Uninspired? Or would you expect randomness, discord, and cacophony? Though artificial intelligence will never be able to duplicate or replace the human composition process, it may be on the way to composing successful music that "breaks enough rules" to be exciting and, perhaps even more importantly, follows enough rules to be listenable! The earliest attempts at using artificial intelligence to compose music were in the 1960s. There are several AI, or machine learning, techniques in use today. The most popular type is called a genetic algorithm. This type of AI mirrors biological evolutionary processes. In it, a "generation" of songs is generated at random or according to predetermined specifications. The best songs are "bred," or spliced together. The process is repeated for many generations, and the songs continually improve. Random mutation is even included as a factor, so that the songs won't prematurely converge and that new ideas are given a chance to thrive. We have to wonder, though: How can the AI know which songs are best? One option is to program in a scoring system with rules for evaluation. These are established and generally-agreed-upon rules of harmonic progressions, voice leading, instrumentation, and rhythm. For instance, unless we're going for a John Cage aesthetic, it's bad to have long periods of silence. Maybe we would also want to specify that if there are multiple parts, they should cross only rarely, if at all. Yet another example would be to say that the tempo should be between 60 and 140 beats per minute. Programmer and composer David Cope from the University of California Santa Cruz has long been involved with a project called "Experiments in Musical Intelligence," in which he guided AI to create new compositions that very convincingly modeled the styles of Mahler, Bach, Beethoven, Chopin, and Joplin. In his case, finding and modeling the rules with which a given composer works was a large part of the challenge. Programming a system of rules has some shortcomings, though, in judging which songs are good enough to be bred for the next generation of songs. As with Cope's work, it may be difficult to even identify the rules. There are a lot of potential rules, and some of them conflict. Finally, the art of composition involves knowing when to follow the rules and when to break them. There's a difficult-to-find balance between the chaos of following no rules and the boringness of strictly following them all. An alternative training method for genetic algorithms that compose music is to have people review the songs to say what's good enough to be bred for the next generation. However, there's a major problem with genetic algorithms, musical and otherwise, where humans are reviewing the quality of the output. It's known as the "fitness bottleneck."  No one wants to listen to thousands of songs, especially at the beginning of the training process when many of the songs might be terrible and near-random. In the current state of research in compositional genetic algorithms, this is one of the big unanswered questions: Should we try to encode extensive sets of rules and rely on them, or find a way to gather feedback from lots of different people so that the workload is shared, or perhaps try a combination of the two that can learn from both rules and human feedback? Crowdsourcing — the term used to describe the collective online intelligence — is a promising solution for the fitness bottleneck problem. Businesses use it to do things that computers aren't very good at, such as tagging images or checking text for profanity. For compositional genetic algorithms, it means that the burden of listening to thousands of developing songs can be spread across a large group of people. In that way, it can be thought of as a large-scale collaborative composition project that doesn't require the participants to have any background knowledge in music. Dr. Bob MacCallum and Armand Leroi from the Imperial College of London developed two related programs called Evolectronica (motto: "survival of the funkiest") and DarwinTunes. These programs take advantage of crowdsourcing to train genetic algorithms and let them know which songs are worth breeding and which aren't. The public has been invited to listen to and rate songs as they evolve. In a similar project, I am using the crowdsourcing marketplace Amazon Mechanical Turk to train a genetic algorithm for my Ph.D. dissertation. Research in compositional AI lets us learn more about how best to implement genetic There's a difficult- to-find balance between the chaos of following no rules and the boringness of strictly following them all. algorithms, which are currently being used to solve problems in many other fields like economics, biology, and engineering. When the genetic algorithm fitness function is crowdsourced, it allows a large group of people to offer their creative influence on the resulting compositions. While we won't replace human composers, we might find some enjoyable and interesting music along the way. Jessica Keup holds a Bachelor's Degree in Information Technology and a Bachelor's Degree in Piano Performance, both from East Tennessee State University, where she currently works as a lecturer. She earned her Master's Degree in Human Computer Interaction at Carnegie Mellon University and is nearing completion of her Ph.D. in Computer Information Systems at Nova Southeastern University. She lives in Jonesborough, TN with her husband Erik and their four cats. Jessica enjoys reading, crafts, all types of music, and being outside in the lovely Appalachian Mountains. She still occasionally performs on piano and has been a church organist for about fourteen years. CLICK FOR MORE To hear MP3s from David Cope's experiments, visit http://artsites.ucsc.edu/faculty/ cope/mp3page.htm. Evolectronica http://evolectronica.com DarwinTunes http://darwintunes.org Amazon Mechanical Turk https://www.mturk.com/ mturk/welcome sai-national.org FALL 2011 PAN PIPES 21

• Cover