Linear Dynamic Programming and the Training of Sequence Estimators

摘要: We consider the problem of finding an optimal path through a trellis graph when arc costs are linear functions unknown parameter vector. In this context we develop algorithm, Linear Dynamic Programming (LDP), that simultaneously computes for all values parameter. show how LDP algorithm can be used supervised learning dynamic-programming-based sequence estimator by minimizing empirical risk. present application to musical harmonic analysis in which optimize performance our seeking value generating best agreeing with hand-labeled data.