Abstract:

GenJam is an interactive genetic algorithm that learns jazz improvisation. It uses two hierarchically related populations to represent melodic ideas at the measure and phrase levels. These populations are evolved using tournament selection, single-point crossover, musically meaningful mutation, and replacement with a 50% generation gap. Fitness for the individual measures and phrases is derived from real-time feedback, which is provided by a human mentor while GenJam improvises to the accompaniment of a synthesized rhythm section. GenJam has been used for actual gigs under the billing Al Biles Virtual Quintet, which features the author on trumpet and GenJam on a variety of synthesized instruments, playing a repertoire of over 90 tunes in a variety of jazz, Latin, and new-age styles. Recent enhancements to GenJam include a pitch-to-MIDI capability that allows GenJam to listen to a human soloist, map his four-bar phrases to the GenJam genetic representation for phrases, apply selected mutation operators to these phrases, and play them back in real time as it trades fours or eights with a human soloist. In this way GenJam is truly interactive in performance, as well as during training. The lecture will feature a live demonstration of GenJam.