Similarity-Based Alignment and Generalization

Abstract

We present a novel approach to learning predictive sequential models, called similarity-based alignment and generalization, which incorporates in the induction process a specific form of domain knowledge derived from a similarity function between the points in the input space. When applied to Hidden Markov Models, our framework yields a new class of learning algorithms called SimAlignGen. We discuss the application of our approach to the problem of programming by demonstration–the problem of learning a procedural model of user behavior by observing the interaction an application Graphical User Interface (GUI). We describe in detail the SimIOHMM, a specific instance of SimAlignGen that extends the known Input-Output Hidden Markov Model (IOHMM). Empirical evaluations of the SimIOHMM show the dependence of the prediction accuracy on the introduced similarity bias, and the computational gains over the IOHMM.