Pattern Matching in David Cope's Experiments in Musical Intelligence
David Cope (1996) reveals that the most successful works in Experiments in Musical Intelligence's (hereafter EMI) output result from a small sample size recombination (typically beat-to-beat). A small sample size requires a large number of recombinancy operations, allowing the output of new works with no direct references to individual works stored in the database. However, signatures are typically larger than one single beat. The pattern matching (unsupervised learning) component in EMI ensures the protection of signatures once these are identified. The identification and protection of signatures is an essential step to preserve the stylistic identity of the database submitted for recombination.
Pattern Matching in Music: Pattern Deviation
Pattern matches in Experiments in Musical Intelligence can operate at several levels including pitch, rhythm, timbre (essential to detect orchestration signatures) and dynamic levels. The matching process is optimized through controllers (variables which control the amount of deviation allowed for a match to occur) (Cope 2003b). The use of controllers enables the identification of patterns within a user defined window of deviation (allowance), thus emulating the human ability to recognize the reoccurrence of musical objects even when these are not exact copies of each other. Slight variations in a recurrent pattern are very common in music, in fact, one should expect to find at least as many approximate as perfect matches. Without a deviation allowance, the pattern-match program would fail to recognize a fugue’s tonal answer as essentially the same as the subject.