Computational models of music composition also often apply an analysis or learning-based approach instead of a rule-based approach. For example, the work of Cope [1996,2000,1991] gained particular interest, also because of the musical quality of his results.
Yet, such an approach is best suited to model an existing style of music for which a corpus of examples is available. Composers, however, are usually less interested in style replication but aim to develop their own distinct musical language. The development of a tool for composers was the original motivation of the present research, which therefore prefers a rule-based approach.
When comparing a rule-based approach with an analysis or learning-based approach, the former expresses explicit musical knowledge (e.g. statements in first-order logic) whereas the musical knowledge for the latter approach is often implicit (e.g. as weights in an artificial neuronal network). However, a learning-based approach can also lead to explicit musical knowledge. For example, [Morales and Morales, 1995] propose a system which automatically creates rules in first-order logic (horn clauses) given a musical example and rule templates. The textbook by Mitchell [1997] introduces learning techniques including the learning of rules.
Rules won by learning can be used in a rule-based system like handwritten rules. Consequently, a generic rule-based system can also be of interest for the community using an analysis or learning-based approach to model music composition.