Luiz Alberto da Silva Abreu

A Hybrid Estimation Scheme Based on the Sequential Importance Resampling Particle Filter and the Particle Swarm Optimization (PSO-SIR)

Particle filters are recursive Bayesian estimators, which are being applied to many areas of engineering in recent years to estimate states and parameters, regarding fire spread, tumors, oil pipelines, heat transfer, chemical reactors, etc. The key idea behind particle filters is that they use an initial distribution (sample), based on the previous state estimate, to calculate the best estimate for the current state, relying only on the current available measurements and the knowledge about the system. The greatest advantage of these methods is the easy computational implementation. However, setting the standard deviation for the initial distribution is very important for the success of the method. For this reason, standard formulation of these methods may not provide good results in problems with large discontinuities (or irregular/abrupt changes). For example, this would be the case of estimating step changes in the heat flux on a plate. Although several solutions have been proposed to improve the estimation performance, they still suffer from the curse of discontinuity. This occurs because particle filters proposed in the literature are not adaptive methods. In the example mentioned above, particle filters can have both a priori information and sample satisfactory before the change. However, after the change begins, the available information could be not enough to draw a suitable sample for the estimation. At this point, it is necessary to modify the standard deviation to broaden the particle search field or to move the a priori information to a new region where a new sample should be drawn. In this regard, the aim of this chapter is to propose a hybrid estimation scheme based on Particle Swarm Optimization (PSO) built into the particle filter Sampling Importance Resampling (SIR) to project the a priori information to a new search region, according to the current observation. To demonstrate the proposal, the problem of estimating step changes on the heat flux on a plate is taken into account, considering experimental measurements. The results allow to state that the scheme combining PSO and SIR provides good performance for this type of problem.

Solução de um Problema Inverso em Dinâmica Populacional através de Inferência Bayesiana e os Métodos Luus-Jaakola e Gauss-Newton.

Neste trabalho e proposto o estudo da dinamica de evolucao de uma populacao na presenca de armadilhas locais. O problema direto e resolvido com a Tecnica de Transformada Integral Generalizada, oferecendo uma solucao de baixo custo computacional. O problema inverso de estimativa de parametros e formulado atraves da abordagem Bayesiana, definindo-se a funcao objetivo de Maximum a Posteriori. A principal contribuicao e a proposta de hibridizacao entre os metodos Luus-Jaakola e Gauss-Newton. Os resultados obtidos mostram que a abordagem e capaz de reduzir a influencia das estimativas iniciais dos parametros.