Data assimilation in models representing spatio-temporal phenomena poses a challenge, particularly if the spatial histogram of the variable appears with multiple modes. The traditional Kalman model is based on a Gaussian initial distribution and Gauss-linear dynamic and observation models. This model is contained in the class of Gaussian distribution and is therefore analytically tractable. It is however unsuitable for representing multimodality. We define the selection Kalman model that is based on a selection-Gaussian initial distribution and Gauss-linear dynamic and observation models. The selection-Gaussian distribution can be seen as a generalization of the Gaussian distribution and may represent multimodality, skewness and peakedness. This selection Kalman model is contained in the class of selection-Gaussian distributions and therefore it is analytically tractable. An efficient recursive algorithm for assessing the selection Kalman model is specified. The synthetic case study of spatio-temporal inversion of an initial state, inspired by pollution monitoring, containing an extreme event suggests that the use of the selection Kalman model offers significant improvements compared to the traditional Kalman model when reconstructing discontinuous initial states.
Combining Ensemble Kalman Filter and Reservoir Computing to predict spatio-temporal chaotic systems from imperfect observations and models
School of Engineering, the University of Tokyo(东京大学工程学院);Meteorological Research Institute, Japan Meteorological Agency(日本气象厅气象研究所);RIKEN Center for Computational Science(理化学研究所计算科学中心)
Recurrent Kalman Networks: Factorized Inference in High-Dimensional Deep Feature Spaces
University of Lincoln(林肯大学);University of Birmingham(伯明翰大学);Lancaster University(兰卡斯特大学);Bosch Center for Artificial Intelligence(博世人工智能中心)
为了将不确定性估计集成到深度时间序列建模中,卡尔曼滤波器(KF)(Kalman et al.,1960)已经与深度学习模型集成,然而,这样的方法通常依赖于近似推理技术,例如变分推理,这使得学习更复杂并且由于近似误差而通常可扩展性较低。我们提出了一种新的深度卡尔曼滤波方法,可以使用反向传播以端到端的方式直接学习,而无需额外的近似。我们的方法使用了高维分解的潜在状态表示,卡尔曼更新简化为标量操作,从而避免了难以反向传播,计算量大,可能不稳定的矩阵求逆。此外,我们使用局部线性动态模型,有效地传播的潜在状态到下一个时间步。由此产生的网络架构,我们称之为递归卡尔曼网络(RKN),可用于任何时间序列数据,类似于LSTM(Hochreiter & Schmidhuber,1997),但使用显式表示不确定性。如我们的实验所示,RKN获得比LSTM或门控递归单元(GRU)更准确的不确定性估计(Cho等人,2014),同时还显示出略微改善的预测性能,并且在图像填补任务上优于各种最近的生成模型。