Path Difference Learning for Guitar Fingering Problem

In this paper we address the problem of mapping guitar music score into one of possible alternative fingering sequences on the fretboard grid. We use dynamic programming (DP) to model the decision process of a guitarist choosing the optimal fingering sequence. To estimate the DP cost functions based on examples of guitar fingering transcriptions (tablatures) we developed an original method named "path difference learning” employing a gradient descent search on the coefficients of the cost function. Features of the fingering alternatives that capture the essence of the mechanical difficulty and musical quality are used to reject impractical fingerings and thus reduce the DP search complexity. Our experiments for several classical guitar pieces showed consistent convergence of the path difference learning. The adaptation resulted in a significant decrease in error count compared to manually selecting the cost function weights.