Two problems with variational expectation maximisation for time-series models

Variational methods are a key component of the approximate inference and learning toolbox. These methods fill an important middle ground, retaining distributional information about uncertainty in latent variables, unlike maximum a posteriori methods (MAP), and yet generally requiring less computational time than Monte Carlo Markov Chain methods. In particular the variational Expectation Maximisation (vEM) and variational Bayes algorithms, both involving variational optimisation of a free-energy, are widely used in time-series modelling. Here, we investigate the success of vEM in simple probabilistic time-series models. First we consider the inference step of vEM, and show that a consequence of the well-known compactness property of variational inference is a failure to propagate uncertainty in time, thus limiting the usefulness of the retained distributional information. In particular, the uncertainty may appear to be smallest precisely when the approximation is poorest. Second, we consider parameter learning and analytically reveal systematic biases in the parameters found by vEM. Surprisingly, simpler variational approximations (such a mean-field) can lead to less bias than more complicated structured approximations.

Dynamic Covariance Models for Multivariate Financial Time Series

The accurate prediction of time-changing covariances is an important problem in the modeling of multivariate financial data. However, some of the most popular models suffer from a) overfitting problems and multiple local optima, b) failure to capture shifts in market conditions and c) large computational costs. To address these problems we introduce a novel dynamic model for time-changing covariances. Over-fitting and local optima are avoided by following a Bayesian approach instead of computing point estimates. Changes in market conditions are captured by assuming a diffusion process in parameter values, and finally computationally efficient and scalable inference is performed using particle filters. Experiments with financial data show excellent performance of the proposed method with respect to current standard models.

Implementation of multi-group cross-section methodology in BGCore MC-depletion code

BGCore is a software package for comprehensive computer simulation of nuclear reactor systems and their fuel cycles. The BGCore interfaces Monte Carlo particles transport code MCNP4C with a SARAF module - an independently developed code for calculating in-core fuel composition and spent fuel emissions following discharge. In BGCore system, depletion coupling methodology is based on the multi-group approach that significantly reduces computation time and allows tracking of large number of nuclides during calculations. In this study, burnup calculation capabilities of BGCore system were validated against well established and verified, computer codes for thermal and fast spectrum lattices. Very good agreement in k eigenvalue and nuclide densities prediction was observed for all cases under consideration. In addition, decay heat prediction capabilities of the BGCore system were benchmarked against the most recent edition of ANS Standard methodology for UO2 fuel decay power prediction in LWRs. It was found that the difference between ANS standard data and that predicted by the BGCore does not exceed 5%.

Multilinear function factorisation for time series feature extraction

This work applies a variety of multilinear function factorisation techniques to extract appropriate features or attributes from high dimensional multivariate time series for classification. Recently, a great deal of work has centred around designing time series classifiers using more and more complex feature extraction and machine learning schemes. This paper argues that complex learners and domain specific feature extraction schemes of this type are not necessarily needed for time series classification, as excellent classification results can be obtained by simply applying a number of existing matrix factorisation or linear projection techniques, which are simple and computationally inexpensive. We highlight this using a geometric separability measure and classification accuracies obtained though experiments on four different high dimensional multivariate time series datasets. © 2013 IEEE.

The forming of variable cross-section I-beams by hot rolling

A process is presented for the forming of variable cross-section I-beams by hot rolling. Optimized I-beams with variable cross-section offer a significant weight advantage over prismatic beams. By tailoring the cross-section to the bending moment experienced within the beam, around 30% of the material can be saved compared to a standard section. Production of such beams by hot rolling would be advantageous, as It combines high volume capacity with high material yields. Through controlled variation of the roll gap during multiple passes, beams with a variable cross-section have been created using shaped rolls similar to those used for conventional I-beam rolling. The process was tested experimentally on a small scale rolling mill, using plasticine as the modelling material. These results were then compared to finite element simulations of individual stages of the process conducted using Abaqus/Standard. Results here show that the process can successfully form a beam with a variable depth web. The main failure modes of the process, and the limitations on the achievable variations In geometry are also presented. Finally, the question of whether or not optimal beam geometries can be created by this process Is discussed. © 2011 Wiley-VCH Verlag GmbH & Co. KGaA. Weinheim.