I've just started working on a statistical approach to generate banjo arrangements in the style of Earl Scruggs, father of bluegrass banjo. Like you said, statistical analyses of existing compositions and recomposition based on that. The difficulty so far is data acquisition & prep - I'm entering & annotating transcriptions manually. A lot of work for something that's unlikely to succeed, but it's fun. If I get any results worth sharing, I'll put it on HN.