Evidential reasoning for adaptive state and parameter estimation in nonlinear systems

点击翻译按钮获取中文摘要

State and parameter estimation is fundamental to control and modeling in many disciplines. Parameter estimation is needed to implement system models that, together with estimates of the current state, are used to formulate intelligent decisions based on sensor information, to compute feedback control signals, to adapt the model or control algorithms and/or parameters, and to build a representation of the operational environment. Classical approaches to state and parameter estimation, such as Kalman or related Bayesian methods, are largely based on linear Gaussian assumptions under which it is sufficient to estimate the mean and covariance of prior and posterior density distributions of the estimates. More recently, new methods have been developed for nonlinear systems based on sequential Monte Carlo particle filtering and its many variants. Particle filters carry more complete information using a point mass representation of probability distributions rather than just the mean and covariance. Our recent research has evaluated particle filtering methods for nonlinear applications and has contributed new techniques to improve performance and robustness of the methods and also to introduce evidential reasoning for representation and accumulation of knowledge based on the Dempter-Shafer theory of evidence. Resources requested under this proposal will fund graduate student research to develop, implement, and test new methods for evidential reasoning in particle filtering for parameter and state estimation. In our research, we will target applications in vehicular robotics, but the results are expected to be applicable to many diverse applications.