International audience

A practical methodology for integrated stochastic replenishment planning with supplier selection is proposed for the single item inventory system. A rolling horizon strategy is adopted to implement the ordering decisions. Our method works in two stages. The first stage is a general black box stage that gives the minimum expected coverage period cost. The second stage uses a dynamic programming approach to compute the minimum expected cost for the rolling horizon. The proposed method is applicable for both stationary and non-stationary demand distributions and even for problems with minimum order quantity constraints. We also propose to examine the benefits of a dynamic supplier selection approach in comparison to selecting a common supplier. We conduct extensive numerical analyses on synthetic data sets for validation.

This article introduces a class of first-order stationary time- varying Pitman-Yor processes. Subsuming our construction of time-varying Dirichlet processes presented in (Caron et al., 2007), these models can be used for time-dynamic density estimation and clustering. Our intuitive and simple construction relies on a generalized Pólya urn scheme. Significantly, this construction yields marginal distributions at each time point that can be explicitly characterized and easily controlled. Inference is performed using Markov chain Monte Carlo and sequential Monte Carlo methods. We demonstrate our models and algorithms on epidemiological and video tracking data.

A nonparametric approach combining generative models and functional data analysis is presented in this paper for classifying functional data which arise naturally in a wide variety of signal processing applications, such as brain computer interfacing, speech recognition, or image classification. Based on a new and improved family of Bayesian classifiers, we extend hierarchical Bayesian classification methodology from vector to functional settings. We provide theoretical and practical motivations to our approach which relies on Dirichlet process mixtures and Gaussian processes. The performance is evaluated on phoneme recognition task, and compared to that of Functional Support Vector Machines (FSVMs).

Choosing the appropriate parameter prior distributions associated to a given bayesian model is a challenging problem. Conjugate priors can be selected for simplicity motivations. However, conjugate priors can be too restrictive to accurately model the available prior information. This paper studies a new generative supervised classifier which assumes that the parameter prior distributions conditioned on each class are mixtures of Dirichlet processes. The motivations for using mixtures of Dirichlet processes is their known ability to model accurately a large class of probability distributions. A Monte Carlo method allowing one to sample according to the resulting class-conditional posterior distributions is then studied. The parameters appearing in the class-conditional densities can then be estimated using these generated samples (following bayesian learning). The proposed supervised classifier is applied to the classification of altimetric waveforms backscattered from different surfaces (oceans, ices, forests, and deserts). This classification is a first step before developing tools allowing for the extraction of useful geophysical information from altimetric waveforms backscattered from nonoceanic surfaces.

This paper presents a method aimed at recognizing environmental sounds for surveillance and security applications. We propose to apply one-class support vector machines (1-SVMs) together with a sophisticated dissimilarity measure in order to address audio classification, and more specifically, sound recognition. We illustrate the performance of this method on an audio database, which consists of 1015 sounds belonging to nine classes. The database used presents high intraclass diversity in temps of signal properties and some kind of interclass similarities. A large discrepancy in the number of items in each class implies nonuniform probability of sound appearances. The method proceeds as follows: first, the use of a set of state-of-the-art audio features is studied. Then, we introduce a set of novel features obtained by combining elementary features. Experiments conducted on a nine-class classification problem show the superiority of this novel sound recognition method. The best recognition accuracy (96.89%) is obtained when combining wavelet-based features, MFCCs, and individual temporal and frequency features. Our 1-SVM-based multiclass classification approach overperforms the conventional hidden Markov model-based system in the experiments conducted, the improvement in the error rate can reach 50%. Besides, we provide empirical results showing that the single-class SVM outperforms a combination of binary SVMs. Additional experiments demonstrate our method is robust to environmental noise.

This paper proposes to apply optimized one-class support vector machines (1-SVMs) as a discriminative framework in order to address a specific audio classification problem. First, since SVM-based classifier with gaussian RBF kernel is sensitive to the kernel width, the width will be scaled in a distribution-dependent way permitting to avoid under-fitting and over-fitting problems. Moreover, an advanced dissimilarity measure will be introduced. We illustrate the performance of these methods on an audio database containing environmental sounds that may be of great importance for surveillance and security applications. The experiments conducted on a multi-class problem show that by choosing adequately the SVM parameters, we can efficiently address a sounds classification problem characterized by complex real-world datasets.

Algorithms involved in applications such as speaker recognition or image classification need to be able to process data which are sets of vectors with variable size. As opposed to the standard setting for kernel methods, where the data are individual vectors, it is difficult to build a reliable reproducing kernel between such sets of unordered vectors. Most effective techniques rely on the design of kernel calculated on densities estimated independently on each set of vectors; however, this calculation can be numerically tricky: therefore these techniques either use poor estimates such as histograms, or assume unjustified restrictive conditions. In this paper, we improve on the existing framework and design kernels between densities, where these are estimated using an effective nonparametric technique, namely the Akaike-Parzen-Rosenblatt (APR) estimate. Closed-form expressions are obtained for positive definite kernels, and simulation results illustrate the soundness of the approach.

This paper presents a Bayesian technique aimed at classifying signals without prior training (clustering). The approach consists of modelling the observed signals, known only through a finite set of samples corrupted by noise, as Gaussian processes. As in many other Bayesian clustering approaches, the clusters are defined thanks to a mixture model. In order to estimate the number of clusters, we assume a priori a countably infinite number of clusters, thanks to a Dirichlet process model over the Gaussian processes parameters. Computations are performed thanks to a dedicated Monte Carlo Markov Chain algorithm, and results involving real signals (mRNA expression profiles) are presented.

This paper addresses the joint estimation and detection of time-varying harmonic components in audio signals. We follow a flexible viewpoint, where several frequency/amplitude trajectories are tracked in spectrogram using particle filtering. The core idea is that each harmonic component (composed of a fundamental partial together with several overtone partials) is considered a target. Tracking requires to define a state-space model with state transition and measurement equations. Particle filtering algorithms rely on a so-called sequential importance distribution, and we show that it can be built on previous multipitch estimation algorithms, so as to yield an even more efficient estimation procedure with established convergence properties. Moreover, as our model captures all the harmonic model information, it actually separates the harmonic sources. Simulations on synthetic and real music data show the interest of our approach

Using Kalman techniques, it is possible to perform optimal estimation in linear Gaussian state-space models. We address here the case where the noise probability density functions are of unknown functional form. A flexible Bayesian nonparametric noise model based on mixture of Dirichlet processes is introduced. Efficient Markov chain Monte Carlo and sequential Monte Carlo methods are then developed to perform optimal estimation in such contexts

Minimum measure sets (MMSs) summarize the information of a (single-class) dataset. In many situations, they can be preferred to estimated probability density functions (pdfs): they are strongly related to pdf level sets while being much easier to estimate in large dimensions. The main contribution of this paper is a theoretical connection between MMSs and one class support vector machines. This justifies the use of one-class SVMs in the following applications: novelty detection (we give explicit convergence rate) and change detection

We propose a joint segmentation algorithm for piecewise constant AR processes recorded by several independent sensors. The algorithm is based on a hierarchical Bayesian model. Appropriate priors allow to introduce correlations between the change locations of the observed signals. Numerical problems inherent to Bayesian inference are solved by a Gibbs sampling strategy. The proposed joint segmentation methodology provides interesting results compared to a signal-by-signal segmentation

We present a sequential Monte Carlo (SMC) method for maximum likelihood (ML) parameter estimation in latent variable models. Standard methods rely on gradient algorithms such as the expectation-maximization (EM) algorithm and its Monte Carlo variants. Our approach is different and motivated by similar considerations to simulated annealing (SA); that is we propose to sample from a sequence of artificial distributions whose support concentrates itself on the set of ML estimates. To achieve this we use SMC methods. We conclude by presenting simulation results on a toy problem and a non-linear non-Gaussian time series model

Ground-penetrating radars (GPRs) are very promising sensors for landmine detection as they are capable of detecting landmines with low metal contents. GPRs deliver so-called Bscan data which are, roughly, vertical slice images of the ground. However, due to the high dielectric permittivity contrast at the air-ground interface, a strong response is recorded at early time by GPRs. This response is the main component of the so-called clutter noise and it blurs the responses of landmines buried at shallow depths. The landmine detection task is therefore quite difficult. This paper proposes a new method for automated detection and localization of buried objects from Bscan records. A support vector machine algorithm for online abrupt change detection is implemented and proves to be efficient in detecting buried landmines from Bscan data. The proposed procedure performance is evaluated using simulated and real data.

The paper investigates a new method for the extraction of spectral trajectories from nonstationary multicomponent narrowband signals. The main idea is to apply statistical filtering so as to track frequency trajectories in the short term Fourier transform. A nonlinear observation model is defined and a special particle filtering algorithm is developed.

The paper presents a fast MCMC (Markov chain Monte Carlo) algorithm specially designed for high dimensional models with block structure. Such models are often met in Bayesian inference problems in signal processing, such as spectral estimation, harmonic analysis, blind deconvolution or signal classification. Our algorithm generates samples distributed according to a posterior distribution. We show that sampling the amplitudes together with the remaining model parameters leads to quicker computations than sampling from the marginal posterior, where amplitudes have been integrated out. Simulation results demonstrate the soundness of this approach for high dimensional models.

We present a study of the statistical behavior of the dissimilarity measure, /spl Dscr//sub 5/ proposed previously (Desobry, F. and Davy, M., Proc. IEEE ICASSP, 2003), and which results from a machine learning-based quantile estimation approach, namely, a single-class support vector machine. This dissimilarity measure possesses the interesting property of being asymptotically equivalent to the Fisher ratio when dealing with radial Gaussian probability density functions. More generally, it can be efficiently applied to non-connected quantiles, and to noisy data sets, as outliers are taken into account by the SVM. A generalisation of /spl Dscr//sub 5/ is then proposed, which results in the design of a more general class of dissimilarity measures, also defined in feature space and with the same properties.

This paper addresses the problem of time-varying amplitude multichirp signals parameter estimation. We compare two approaches which require a model for the amplitude. First, we use a basis of time-localized functions associated with Bayesian estimation. Secondly, we use an autoregressive model associated with a mixed high-order ambiguity function/Kalman filter estimation. Results show that both methods are efficient to solve this estimation problem.

We present a machine learning technique aimed at detecting abrupt changes in a sequence of vectors. Our algorithm requires a Mercer kernel together with the corresponding feature space. A stationarity index is designed in the feature space, and consists of comparing two circles corresponding to two /spl nu/-SV novelty detectors via a Fisher-like ratio. An abrupt change corresponds to a large distance between the circle centers (with respect to their radii). We show that the index can be computed in the input space, and simulation results show its efficiency in front of real data.

We establish connections between Cohen-Posch (1985) theory of positive time-frequency distributions (TFDs) and copula theory. Both are aimed at designing joint probability distributions with fixed marginals, and we demonstrate that they are formally equivalent. Moreover, we show that copula theory leads to a noniterative method for constructing positive TFDs. Simulations show typical results.

We address here the problem of developing efficient particle filtering techniques in order to estimate the state of Jump Markov Systems (JMS). These processes are often met in signal processing (target tracking, communication…). Our algorithm takes advantage of the structure of the process. We apply our algorithm to time varying autoregressive processes.

In this paper, we introduce an hybrid time-frequency/support vector machine algorithm for the detection of abrupt spectral changes. A stationarity index is derived from support vector novelty detection theory by using sub-images extracted from the time-frequency plane as feature vectors. Simulations show the efficiency of this new algorithm for audio signal segmentation, compared to another nonparametric detector.

This paper addresses the problem of classifying chirp signals using hierarchical Bayesian learning combined with Markov Chain Monte Carlo (MCMC) methods. Bayesian learning consists of estimating the distribution of observed data conditional upon each class from a set of training samples. Unfortunately, this estimation often requires to evaluate intractable multidimensional integrals. This paper studies an original implementation of hierarchical Bayesian learning which estimates the class conditional probability densities using MCMC methods. The performance of this implementation is compared to other existing approaches for the classification of chirp signals.

Estimating the pitch of musical signals is complicated by the presence of partials in addition to the fundamental frequency. In this paper, we propose developments to an earlier Bayesian model which describes each component signal in terms of fundamental frequency, partials (‘harmonics’), and amplitude. This basic model is modified for greater realism to include non-white residual spectrum, time-varying amplitudes and partials ‘detuned’ from the natural linear relationship. The unknown parameters of the new model are simulated using a reversible jump MCMC algorithm, leading to a highly accurate pitch estimator. The models and algorithms can be applied for feature extraction, polyphonic music transcription, source separation and restoration of musical sources.

This letter describes an efficient method to perform nonstationary signal classification. A support vector machine (SVM) algorithm is introduced and its parameters optimized in a principled way. Simulations demonstrate that our low-complexity method outperforms state-of-the-art nonstationary signal classification techniques.

This paper addresses the problem of loudspeaker testing. Two applications are investigated: loudspeaker manufacturing fault detection and maintenance. In order to comply with practical requirements, we propose a new, brief, nonstationary test signal. Recorded loudspeaker responses are processed using an improved time-frequency decision algorithm. The fault detection procedure is tested with real data. The results show its accuracy and practical interest.

Time-frequency representations (TFRs) are efficient tools for nonstationary signal classification. However, the choice of the TFR and of the distance measure employed is critical when no prior information other than a learning set of limited size is available. In this letter, we propose to jointly optimize the TFR and distance measure by minimizing the (estimated) probability of classification error. The resulting optimized classification method is applied to multicomponent chirp signals and real speech records (speaker recognition). Extensive simulations show the substantial improvement of classification performance obtained with our optimization method.

This paper addresses the problem of supervised classification using general Bayesian learning. General Bayesian learning consists of estimating the unknown class-conditional densities from a set of labelled samples. However, the estimation requires to evaluate intractable multidimensional integrals. This paper studies an implementation of general Bayesian learning based on Markov chain Monte Carlo (MCMC) methods.