A new particle filter for high-dimensional state-space models based on intensive and extensive proposal distribution

Abstract

This paper proposes novel principles and techniques of a particle filter to estimate dynamic system states under an observed time series data and a state-space model which are possibly non-linear and have the dimensions more than several hundreds. First, we point out two crucial curses of dimensionality and propose three key ideas to overcome them. Second, we propose the novel particle filters implementing these ideas and analyse their mathematical characteristics. Our experimental evaluation demonstrates their significant accuracy, robustness and efficiency for both artificial and real-world problems having large scales.