Tuned iterated filtering

Abstract

Iterated filtering is an algorithm for estimating parameters in partially observed Markov process (POMP) models. The real-world performance of the algorithm depends on several tuning parameters. We propose a simple method for optimizing the parameter governing the joint dynamics of the hidden parameter process (called the Σ matrix).The tuning is implemented using a fixed-lag sequential Monte Carlo expectation–maximization (EM) algorithm. We introduce two different versions of the tuning parameter, the approximately estimated Σ matrix, and a normalized version of the same matrix.Our simulations show that the finite-sample performance for the normalized matrix outperform the standard iterated filter, while the naive version is doing more harm than good.