Training Hidden Markov Models using Population-Based Learning

Hidden Markov Models are commonly trained using algorithms derived from gradient-based methods such as the Baum-Welch procedure. We describe a new representation of discrete observation HMMs that permits them to be trained using Population-Based Incremental Learning (PBIL), a variant of genetic learning that combines evolutionary optimization and hill-climbing Baluja and Caruana, 1995]. In this paper we examine the recognition performance of PBIL-trained HMMs on two tasks: hand-drawn shape recognition and spoken digit recognition. We demonstrate that HMMs can be maximized via PBIL using either a maximum likelihood or a maximum mutual information tness function and achieve results comparable to those achieved using the Baum-Welch procedure. We argue that the PBIL algorithm has the advantage of easy extension to applications like speech intelligibility assessment that lack a diierentiable or analytical optimization function.