Abstract: This paper deals with the problem of estimating the volume of the excursion set of a function $f:\mathbb{R}^d \to \mathbb{R}$ above a given threshold, under a probability measure on $\mathbb{R}^d$ that is assumed to be known. In the industrial world, this corresponds to the problem of estimating a probability of failure of a system. When only an expensive-to-simulate model of the system is available, the budget for simulations is usually severely limited and therefore classical Monte Carlo methods ought to be avoided. One of the main contributions of this article is to derive SUR (stepwise uncertainty reduction) strategies from a Bayesian-theoretic formulation of the problem of estimating a probability of failure. These sequential strategies use a Gaussian process model of $f$ and aim at performing evaluations of $f$ as efficiently as possible to infer the value of the probability of failure. We compare these strategies to other strategies also based on a Gaussian process model for estimating a probability of failure.
Abstract: This paper deals with the convergence of the expected improvement algorithm, a popular global optimization algorithm based on a Gaussian process model of the function to be optimized. The first result is that under some mild hypotheses on the covariance function $k$ of the Gaussian process, the expected improvement algorithm produces a dense sequence of evaluation points in the search domain, when the function to be optimized is in the reproducing kernel Hilbert space (RKHS) generated by $k$. The second result states that the density property also holds for $\P$-almost all continuous functions, where $\P$ is the (prior) probability distribution induced by the Gaussian process.
Abstract: A general formulation of the Fokker–Planck–Kolmogorov (FPK) equation for stochastic hybrid systems is presented, within the framework of Generalized Stochastic Hybrid Systems (GSHSs). The FPK equation describes the time evolution of the probability law of the hybrid state. Our derivation is based on the concept of mean jump intensity, which is related to both the usual stochastic intensity (in the case of spontaneous jumps) and the notion of probability current (in the case of forced jumps). This work unifies all previously known instances of the FPK equation for stochastic hybrid systems, and provides GSHS practitioners with a tool to derive the correct evolution equation for the probability law of the state in any given example.
Abstract: We consider the problem of optimizing a real-valued continuous function $f$ using a Bayesian approach, where the evaluations of $f$ are chosen sequentially by combining prior information about $f$, which is described by a random process model, and past evaluation results. The main difficulty with this approach is to be able to compute the posterior distributions of quantities of interest which are used to choose evaluation points. In this article, we decide to use a Sequential Monte Carlo (SMC) approach.
Abstract: We consider the problem of optimizing a real-valued continuous function f, which is supposed to be expensive to evaluate and, consequently, can only be evaluated a limited number of times. This article focuses on the Bayesian approach to this problem, which consists in combining evaluation results and prior information about f in order to efficiently select new evaluation points, as long as the budget for evaluations is not exhausted.
The algorithm called efficient global optimization (EGO), proposed by Jones, Schonlau and Welch (J. Global Optim., 13(4):455–492, 1998), is one of the most popular Bayesian optimization algorithms. It is based on a sampling criterion called the expected improvement (EI), which
assumes a Gaussian process prior about f. In the EGO algorithm, the parameters of the covariance of the Gaussian process are estimated from the evaluation results by maximum likelihood, and these parameters are then plugged in the EI sampling criterion. However, it is well-known that this plug-in strategy can lead to very disappointing results when the evaluation results do not carry enough information about f to estimate the parameters in a satisfactory manner.
We advocate a fully Bayesian approach to this problem, and derive an analytical expression for the EI criterion in the case of Student predictive distributions. Numerical experiments show that the fully Bayesian approach makes EI-based optimization more robust while maintaining an average loss similar to that of the EGO algorithm.
Abstract: This paper addresses the problem of summarizing the posterior distributions that typically arise, in a Bayesian framework, when dealing with signal decomposition problems with unknown number of components. Such posterior distributions are defined over union of subspaces of differing dimensionality and can be sampled from using modern Monte Carlo techniques, for instance the increasingly popular RJ-MCMC method. No generic approach is available, however, to summarize the resulting variable-dimensional samples and extract from them component-specific parameters.
We propose a novel approach to this problem, which consists in approximating the complex posterior of interest by a "simple"---but still variable-dimensional---parametric distribution. The distance between the two distributions is measured using the Kullback-Leibler divergence, and a Stochastic EM-type algorithm, driven by the RJ-MCMC sampler, is proposed to estimate the parameters. The proposed algorithm is illustrated on the fundamental signal processing example of joint detection and estimation of sinusoids in white Gaussian noise.
Abstract: This paper addresses the behavior in low SNR situations of the algorithm proposed by Andrieu and Doucet (IEEE T. Signal Proces., 47(10), 1999) for the joint Bayesian model selection and estimation of sinusoids in Gaussian white noise. It is shown that the value of a certain hyperparameter, claimed to be weakly influential in the original paper, becomes in fact quite important in this context. This robustness issue is fixed by a suitable modification of the prior distribution, based on model selection considerations. Numerical experiments show that the resulting algorithm is more robust to the value of its hyperparameters.
Abstract: This paper deals with several issues related to the pointwise consistency of the kriging predictor when the mean and the covariance functions are known. These questions are of general importance in the context of computer experiments. The analysis is based on the properties of approximations in reproducing kernel Hilbert spaces. We fix an erroneous claim of Yakowitz and Szidarovszky (J. Multivariate Analysis, 1985) that the kriging predictor is pointwise consistent for all continuous sample paths under some assumptions.
Abstract: This paper deals with the problem of estimating the probability of failure of a system, in the challenging case where only an expensive-to-simulate model is available. In this context, the budget for simulations is usually severely limited and therefore classical Monte~Carlo methods ought to be avoided. We present a new strategy to address this problem, in the framework of sequential Bayesian planning. The method uses kriging to compute an approximation of the probability of failure, and selects the next simulation to be conducted so as to reduce the mean square error of estimation. By way of illustration, we estimate the probability of failure of a control strategy in the presence of uncertainty about the parameters of the plant.
Abstract: A general formulation of the Fokker-Planck-Kolmogorov (FPK) equation for stochastic hybrid systems is presented, within the framework of Generalized Stochastic Hybrid Systems (GSHS). The FPK equation describes the time evolution of the probability law of the hybrid state. Our derivation is based on the concept of mean jump intensity, which is related to both the usual stochastic intensity (in the case of spontaneous jumps) and the notion of probability current (in the case of forced jumps). This work unifies all previously known instances of the FPK equation for stochastic hybrid systems, and provides GSHS practitioners with a tool to derive the correct evolution equation for the probability law of the state in any given example.
Abstract: This paper addresses the problem of predicting a wind farm's power generation when no or few statistical data is available. The study is based on a time-series wind speed model and on a simple dynamic model of a DFIG wind turbine including cut-off and cut-in behaviours. The wind turbine is modeled as a stochastic hybrid system with three operation modes. Numerical results, obtained using Monte-Carlo simulations, provide the annual distribution of a wind farm's active power generation. For different numbers of wind turbines, we compare the numerical results obtained using the dynamic model with those obtained considering the wind turbine's steady-state power curve. Simulations show that the wind turbine's dynamics do not need to be considered for analyzing the annual distribution of a wind farm generation.
Abstract: This paper presents some recent results concerning a class of continuous-time Markov processes called ``stochastic hybrid systems''. These processes describe the evolution of a multidimensional hybrid-state dynamical system subject to Gaussian white noise inputs. After a brief recall of the formalism, we state the generalized Fokker-Planck equation, which is a partial differential equation satisfied by the probability density function of the system. As an illustration, we consider a variable-speed wind turbine, with a switching controller that combines stall regulation and pitch control. For a given value of the mean wind speed, the stationary distribution of the state variables is computed numerically. This truly dynamical analysis of the system yields a complete probabilistic characterization of the uncertain power output, which is much more accurate than the usual static analysis.
Abstract: Among image restoration literature, there are mainly two kinds of approach. One is based on a process over image wavelet coefficients, as wavelet shrinkage for denoising. The other one is based on a process over image gradient. In order to get an edge-preserving regularization, one usually assume that the image belongs to the space of functions of Bounded Variation (BV). An energy is minimized, composed of an observation term and the Total Variation (TV) of the image. Recent contributions try to mix both types of method. In this spirit, the goal of this paper is to define a unified framework including together wavelet methods and energy minimization as TV. In fact, for denoising purpose, it is already shown that wavelet soft-thresholding is equivalent to choose the regularization term as the norm of the Besov space $B^{11}_1$. In the present work, this equivalence result is extended to the case of deconvolution problem. We propose a general functional to minimize, which includes the TV minimization, wavelet coefficients regularization, mixed (TV+wavelet) regularization or more general terms. Moreover we give a projection-based algorithm to compute the solution. The convergence of the algorithm is also stated. We show that the decomposition of an image over a dictionary of elementary shapes (atoms) is also included in the proposed framework. So we give a new algorithm to solve this difficult problem, known as Basis Pursuit. We also show numerical results of image deconvolution using TV, wavelets, or TV+wavelets regularization terms.
Abstract: A Stepwise Uncertainty Reduction (SUR) strategy aims at constructing a sequence X1(f),X2(f), . . . of evaluation points of a function f : Rd → R in such a way that the residual uncertainty about a quantity of interest S(f) given the information provided by the evaluation results is small. In Bect, Ginsbourger, Li, Picheny and Vazquez, Statistics and Computing, 2011, several SUR approaches have been shown to be particularly efficient for the problem of estimating the volume of an excursion set of a function f above a threshold. Here, we build upon these results and we present fast implementations of some SUR strategies, which are based on two ideas. The first idea is to take advantage of update formulas for kriging. The second idea is to derive closed-form expressions for some integrals that appear in the SUR criteria. We are able to demonstrate significant speed-ups and we illustrate our algorithms on a nuclear safety application.
Abstract: Reversible jump MCMC (RJ-MCMC) sampling techniques, which allow to jointly tackle model selection and parameter estimation problems in a coherent Bayesian framework, have become increasingly popular in the signal processing literature since the seminal paper of Andrieu and Doucet (IEEE Trans. Signal Process., 47(10), 1999). Crucial to the implementation of any RJ-MCMC sampler is the computation of the so-called Metropolis-Hastings-Green (MHG) ratio, which determines the acceptance probability for the proposed moves.
This note discusses the computation of the MHG ratio, in the case where the proposal kernel can be decomposed as a mixture of simpler kernels, for which the MHG ratio is easy to compute. We provide sufficient conditions under which the MHG ratio of the mixture can be deduced from the MHG ratios of the elementary kernels of which it is composed.
As an application, we consider the case of Birth-or-Death moves---the simplest kind of trans-dimensional move, which is used in virtually all applications of RJ-MCMC to signal decomposition problems. It turns out that the expression of the MHG ratio that was given in the paper of Andrieu and Doucet was erroneous. Unfortunately, this mistake has contaminated most subsequent papers dealing with RJ-MCMC sampling in the signal processing literature.
Abstract: We consider a Markov process on a Riemannian manifold, which solves a stochastic differential equation in the interior of the manifold and jumps according to a deterministic reset map when it reaches the boundary. We derive a partial differential equation for the probability density function, involving a non-local boundary condition which accounts for the jumping behaviour of the process. This is a generalisation of the usual Fokker-Planck-Kolmogorov equation for diffusion processes. The result is illustrated with an example in the field of stochastic hybrid systems.
