Bayesian optimization for parameter estimation of local particle filter

抄録

The particle filter (PF) is a powerful data assimilation method that does not assume linearity or Gaussianity. However, its application to numerical weather prediction is limited by the exponentially increasing number of particles required as the dimensionality of the dynamical system rises. Although a local particle filter (LPF) achieves the PF in high-dimensional systems through localization, the LPF remains unstable owing to its high parameter sensitivity. In the PF, maintaining particle diversity is essential to prevent “weight collapse,” and an inflation factor that smooths weights among particles is a crucial parameter that should be optimized in the LPF.

Bayesian optimization (BO) is a method for parameter estimation that minimizes (or maximizes) an objective function and is used for parameter optimization of neural networks. This study discussed the benefits of using BO within the LPF framework. As a proof of concept, we used BO to estimate the inflation factor that minimizes the root mean square error between observations and forecasts in the Lorenz-96 40-variable model. The BO quickly estimated the optimal inflation factor equivalent to the brute-force method, allowing the LPF to work stably for a decade scale. Furthermore, our method demonstrated robustness to changes in initial conditions and observations. In conclusion, using BO could greatly reduce the burden and computational cost associated with parameter optimization. The development of BO is expected to lead to further practical application of the LPF and ultimately improve the accuracy of forecasts for torrential rainfall. The benefits of BO will eventually be demonstrated in experiments with atmospheric models aimed at the practical application of the LPF.