Aggregation-based model reduction of a Hidden Markov Model

This paper is concerned with developing an information-theoretic framework to aggregate the state space of a Hidden Markov Model (HMM) on discrete state and observation spaces. The optimal aggregation is obtained by minimizing the Kullback-Leibler (K-L) divergence rate between joint laws describing the state and observation processes. The solution to this optimization problem is just the optimal aggregated Hidden Markov Model. This optimization problem is solved in two steps: The first step is to formulate the optimal solution for any fixed partition. The second step is to find the optimal partition by using an approximate dynamic programming framework. The algorithm can be implemented using a single sample path of the HMM. The convergence of the algorithm is established using standard stochastic approximation arguments and illustrated with the aid of examples.

• Type: Journal
• Year: 2010
• Where: CDC
• Authors: Kun Deng, Prashant G. Mehta, Sean P. Meyn
• Keywords: CDC 2010, Control Systems, Hidden Markov Model, Optimal Aggregation, Optimization Problem