Modelling sequential data is important in many areas of science and engineering. Hidden Markov models (HMMs) and Kalman filter models (KFMs) are popular for this because they are simple and flexible. For example, HMMs have been used for speech recognition and bio-sequence analysis, and KFMs have been used for problems ranging from tracking planes and missiles to predicting the economy. However, HMMs and KFMs are limited in their “expressive power”. Dynamic Bayesian Networks (DBNs) generalize HMMs by allowing the state space to be represented in factored form, instead of as a single discrete random variable. DBNs generalize KFMs by allowing arbitrary probability distributions, not just (unimodal) linear-Gaussian. In this thesis, I will discuss how to represent many different kinds of models as DBNs, how to perform exact and approximate inference in DBNs, and how to learn DBN models from sequential data. In particular, the main novel technical contributions of this thesis are as follows: a way of representing Hierarchical HMMs as DBNs, which enables inference to be done in O(T) time instead of O(T 3), where T is the length of the sequence; an exact smoothing algorithm that takes O(log T) space instead of O(T); a simple way of using the junction tree algorithm for online inference in DBNs; new complexity bounds on exact online inference in DBNs; a new deterministic approximate inference algorithm called factored frontier; an analysis of the relationship between the BK algorithm and loopy belief propagation; a way of applying Rao-Blackwellised particle filtering to DBNs in general, and the SLAM (simultaneous localization and mapping) problem in particular; a way of extending the structural EM algorithm to DBNs; and a variety of different applications of DBNs. However, perhaps the main value of the thesis is its catholic presentation of the field of sequential data modelling.

This paper analyses the recently suggested particle approach to filtering time series. We suggest that the algorithm is not robust to outliers for two reasons: the design of the simulators and the use of the discrete support to represent the sequentially updating prior distribution. Both problems are tackled in this paper. We believe we have largely solved the first problem and have reduced the order of magnitude of the second. In addition we introduce the idea of stratification into the particle filter which allows us to perform on-line Bayesian calculations about the parameters which index the models and maximum likelihood estimation. The new methods are illustrated by using a stochastic volatility model and a time series model of angles. Some key words: Filtering, Markov chain Monte Carlo, Particle filter, Simulation, SIR, State space. 1 1

Bayesian networks are directed acyclic graphs that represent dependencies between variables in a probabilistic model. Many time series models, including the hidden Markov models (HMMs) used in speech recognition and Kalman filter models used in filtering and control applications, can be viewed as examples of dynamic Bayesian networks. We first provide a brief tutorial on learning and Bayesian networks. We then present some dynamic Bayesian networks that can capture much richer structure than HMMs and Kalman filters, including spatial and temporal multiresolution structure, distributed hidden state representations, and multiple switching linear regimes. While exact probabilistic inference is intractable in these networks, one can obtain tractable variational approximations which call as subroutines the forward-backward and Kalman filter recursions. These approximations can be used to learn the model parameters...

We present a probabilistic generarive model for timing deviations in expressive music performance. The structure of the proposed model is equivalent to a switching state space model. The switch variables correspond to discrete note locations as in a musical score. The continuous hidden variables denote the tempo. We formulate two well known music recognition problems, namely tempo tracking and automatic transcription (rhythm quantization) as filtering and maximum a posteriori (MAP) state estimation tasks. Ex- act computation of posterior features such as the MAP state is intractable in this model class, so we introduce Monte Carlo methods for integration and optimization. We compare Markov Chain Monte Carlo (MCMC) methods (such as Gibbs sampling, simulated annealing and iterative improvement) and sequential Monte Carlo methods (particle filters). Our simulation results suggest better results with sequential methods. The methods can be applied in both online and batch scenarios such as tempo tracking and transcription and are thus potentially useful in a number of music applications such as adaptive automatic accompaniment, score typesetting and music information retrieval.

We show how many different variants of Switching Kalman Filter models can be represented in a unified way, leading to a single, general-purpose inference algorithm. We then show how to find approximate Maximum Likelihood Estimates of the parameters using the EM algorithm, extending previous results on learning using EM in the non-switching case [DRO93, GH96a] and in the switching, but fully observed, case [Ham90]. 1 Introduction Dynamical systems are often assumed to be linear and subject to Gaussian noise. This model, called the Linear Dynamical System (LDS) model, can be defined as x t = A t x t\Gamma1 + v t y t = C t x t +w t where x t is the hidden state variable at time t, y t is the observation at time t, and v t ¸ N(0; Q t ) and w t ¸ N(0; R t ) are independent Gaussian noise sources. Typically the parameters of the model \Theta = f(A t ; C t ; Q t ; R t )g are assumed to be time-invariant, so that they can be estimated from data using e.g., EM [GH96a]. One of the main adva...

Modelling and reconstruction methods are presented for noise reduction of autocorrelated signals in non-Gaussian, impulsive noise environments. A Bayesian probabilistic framework is adopted and Markov chain Monte Carlo methods are developed for detection and correction of impulses. Individual noise sources are modelled as Gaussian with unknown scale (variance), allowing for robustness to `heavy-tailed' impulse distributions, while the underlying signal is modelled as autoregressive (AR). Results are presented for both artificial and real data from voice and music recordings and comparisons are made with existing techniques. The new techniques are found to give improved detection and elimination of impulses in adverse noise conditions at the expense of some extra computational complexity.

Abstract not found

This article considers Markov chain computational methods for incorporating uncertainty about the dimension of a parameter when performing inference within a Bayesian setting. A general class of methods is proposed for performing such computations, based upon a product space representation of the problem which is similar to that of Carlin and Chib. It is shown that all of the existing algorithms for incorporation of model uncertainty into Markov chain Monte Carlo (MCMC) can be derived as special cases of this general class of methods. In particular, we show that the popular reversible jump method is obtained when a special form of Metropolis--Hastings (M--H) algorithm is applied to the product space. Furthermore, the Gibbs sampling method and the variable selection method are shown to derive straightforwardly from the general framework. We believe that these new relationships between methods, which were until now seen as diverse procedures, are an important aid to the understanding of MCMC model selection procedures and may assist in the future development of improved procedures. Our discussion also sheds some light upon the important issues of "pseudo-prior" selection in the case of the Carlin and Chib sampler and choice of proposal distribution in the case of reversible jump. Finally, we propose efficient reversible jump proposal schemes that take advantage of any analytic structure that may be present in the model. These proposal schemes are compared with a standard reversible jump scheme for the problem of model order uncertainty in autoregressive time series, demonstrating the improvements which can be achieved through careful choice of proposals

This paper provides a Bayesian analysis of such a model. The main contribution of our paper is that different features of the data--such as the spectral density of the stationary term, the regression parameters, unknown frequencies and missing observations--are combined in a hierarchical Bayesian framework and estimated simultaneously. A Bayesian test to detect the presence of deterministic components in the data is also constructed. Applications of our methods to simulated and real data suggest that they perform well. We place a smoothness prior, similar to that in Wahba (1980), on the logarithm of the spectral density. To make the estimation of the spectral density computationally tractable, Whittle's (1957) approximation to the Gaussian likelihood is used. This results in a nonparametric regression problem with the logarithm of the periodogram as the dependent variable, the logarithm of the spectral density as the unknown regression curve, and observation errors having log chi-squared distributions. By approximating the logarithm of a chi-squared distribution as a mixture of normals, the approximate log likelihood together with the prior for the spectral density can be expressed as a state space model with errors that are mixtures of normals. The computation is carried out efficiently by Markov chain Monte Carlo using the sampling approach in Carter and Kohn (1994). To make the paper easier to read the full model is introduced in a number of steps. Section 2 shows how to estimate the spectral density of a stationary process in the absence of deterministic components. Section 3 extends the estimation to the signal plus noise model with missing observations. Section 4 shows by example how the results in Sections 2 and 3 can be combined to analyze data and studies emp...

Particle filters are sequential Monte Carlo methods that are used in numerous problems where time-varying signals must be presented in real time and where the objective is to estimate various unknowns of the signal and/or detect events described by the signals. The standard solutions of such problems in many applications are based on the Kalman filters or extended Kalman filters. In situations when the problems are nonlinear or the noise that distorts the signals is non-Gaussian, the Kalman filters provide a solution that may be far from optimal. Particle filters are an intriguing alternative to the Kalman filters due to their excellent performance in very di#cult problems including communications, signal processing, navigation, and computer vision. Hence, particle filters have been the focus of wide research recently and immense literature can be found on their theory. Most of these works recognize the complexity and computational intensity of these filters, but there has been no e#ort directed toward the implementation of these filters in hardware. The objective of this dissertation is to develop, design, and build e#cient hardware for particle filters, and thereby bring them closer to practical applications. The fact that particle filters outperform most of the traditional filtering methods in many complex practical scenarios, coupled with the challenges related to decreasing their computational complexity and improving real-time performance, makes this work worthwhile. The main