A0794
Title: Application of data assimilation to particle simulation and point process
Authors: Kazuyuki Nakamura - Meiji University (Japan) [presenting]
Noriho Fujioka - Meiji University (Japan)
Yutaka Kono - Meiji University (Japan)
Abstract: Data assimilation is a method which combines a physical simulation system with observation data to obtain more precise simulation and prediction results. It has been developed in meteorology and oceanography, and recently its application fields are widely spreading such as geotechnical engineering, epidemiology, ecology and molecular dynamics. Data assimilation can be interpreted as Bayesian calculation and sequential Bayesian filtering, and therefore appropriate modeling of the prior and the system noise is important for obtaining suitable estimation. We show the rule for prior design, especially focusing on the constraint conditions for the real systems. Based on the above, we introduce several approaches of prior design for data assimilation of particle simulation and point process such as traffic flow and event occurrence. One of the approaches is the use of bounded Gaussian and uniform mixture, which can manage both fast estimation in sudden parameter change and estimation stability. The validity of the approach is also shown by numerical experiments.
