A0340
Title: Learning algorithms for constrained hidden Markov models
Authors: Connor Mattes - Sandia National Laboratory (United States) [presenting]
William Hart - Sandia National Laboratory (United States)
Abstract: Hidden Markov models (HMMs) are an invaluable tool for modeling discrete processes with hidden (latent) variables. HMMs have proven useful in fields ranging from speech recognition and natural language processing to bioinformatics and bird migration. The focus is specifically on constrained HMMs. Constrained HMMs capture subject matter expertise by restricting the set of hidden variables. For example, hidden states may incur costs that must be bounded, or it may require that some hidden state occur at least once. Although there has been some prior research on learning for constrained HMMs there are significant gaps in the literature. Learning in constrained HMMs is difficult because it is computationally intractable to sum over all feasible hidden states. Towards this end, three novel learning algorithms are introduced that efficiently approximate model parameters by generating many high-quality feasible sets of hidden states. The advantages and disadvantages of each of these new algorithms are discussed. Those are compared theoretically and computationally to the existing literature.
